{
  "schemaVersion": 1,
  "deviceProperties": [
    {
      "id": 0, "name": "AMD Instinct MI300X VF", "totalGlobalMem": 205571227648,
      "computeMajor": 9, "computeMinor": 4,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 64,
      "sharedMemPerBlock": 65536, "numSms": 304
    , "maxSharedMemoryPerMultiProcessor": 65536
    }
  ],
  "roctracer_version": 4.1,
  "hip_runtime_version": 60342131,
  "hip_driver_version": 60342131,
  "with_stack": 1,
  "trace_id": "2ADDA7173CFF47C1A417F679EC4527DC",
  "traceEvents": [
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::embedding", "pid": 537, "tid": 537,
    "ts": 7429564820781.066, "dur": 97.237,
    "args": {
      "External id": 1,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 291, "finished": true, "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564820809.006, "dur": 6.928,
    "args": {
      "External id": 2,"Record function id": 0, "finished": true, "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564820810.382, "dur": 3.859,
    "args": {
      "External id": 3,"Record function id": 0, "finished": true, "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::index_select", "pid": 537, "tid": 537,
    "ts": 7429564820817.469, "dur": 55.629,
    "args": {
      "External id": 4,"Record function id": 0, "finished": true, "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564820829.578, "dur": 6.580,
    "args": {
      "External id": 5,"Record function id": 0, "finished": true, "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 537, "tid": 537,
    "ts": 7429564820840.450, "dur": 5.581,
    "args": {
      "External id": 6,"Record function id": 0, "finished": true, "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564820873.754, "dur": 0.619,
    "args": {
      "External id": 7,"Record function id": 0, "finished": true, "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 537, "tid": 537,
    "ts": 7429564820904.913, "dur": 16.103,
    "args": {
      "External id": 8,"Record function id": 0, "finished": true, "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564820906.485, "dur": 1.122,
    "args": {
      "External id": 9,"Record function id": 0, "finished": true, "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 537, "tid": 537,
    "ts": 7429564820908.504, "dur": 12.172,
    "args": {
      "External id": 10,"Record function id": 0, "finished": true, "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 537, "tid": 537,
    "ts": 7429564820909.886, "dur": 2.108,
    "args": {
      "External id": 11,"Record function id": 0, "finished": true, "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564820930.236, "dur": 10.593,
    "args": {
      "External id": 12,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 11
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564820937.722, "dur": 1.280,
    "args": {
      "External id": 13,"Record function id": 0, "finished": true, "Ev Idx": 12
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::full", "pid": 537, "tid": 537,
    "ts": 7429564820977.888, "dur": 18.724,
    "args": {
      "External id": 14,"Record function id": 0, "finished": true, "Ev Idx": 13
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564820978.990, "dur": 2.217,
    "args": {
      "External id": 15,"Record function id": 0, "finished": true, "Ev Idx": 14
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 537, "tid": 537,
    "ts": 7429564820982.326, "dur": 13.957,
    "args": {
      "External id": 16,"Record function id": 0, "finished": true, "Ev Idx": 15
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::triu", "pid": 537, "tid": 537,
    "ts": 7429564821001.122, "dur": 12.711,
    "args": {
      "External id": 17,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 16
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 537, "tid": 537,
    "ts": 7429564821018.955, "dur": 9.243,
    "args": {
      "External id": 18,"Record function id": 0, "finished": true, "Ev Idx": 17
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564821019.744, "dur": 0.643,
    "args": {
      "External id": 19,"Record function id": 0, "finished": true, "Ev Idx": 18
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 537, "tid": 537,
    "ts": 7429564821020.729, "dur": 7.250,
    "args": {
      "External id": 20,"Record function id": 0, "finished": true, "Ev Idx": 19
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 537, "tid": 537,
    "ts": 7429564821021.176, "dur": 1.229,
    "args": {
      "External id": 21,"Record function id": 0, "finished": true, "Ev Idx": 20
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564821032.064, "dur": 2.883,
    "args": {
      "External id": 22,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 21
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564821032.653, "dur": 2.153,
    "args": {
      "External id": 23,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 22
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::gt", "pid": 537, "tid": 537,
    "ts": 7429564821037.516, "dur": 14.371,
    "args": {
      "External id": 24,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 23
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul_", "pid": 537, "tid": 537,
    "ts": 7429564821055.610, "dur": 13.517,
    "args": {
      "External id": 25,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 24
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564821074.153, "dur": 2.790,
    "args": {
      "External id": 26,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 25
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821075.124, "dur": 1.233,
    "args": {
      "External id": 27,"Record function id": 0, "finished": true, "Ev Idx": 26
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564821077.497, "dur": 2.247,
    "args": {
      "External id": 28,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 27
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821078.208, "dur": 0.827,
    "args": {
      "External id": 29,"Record function id": 0, "finished": true, "Ev Idx": 28
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821086.474, "dur": 3.920,
    "args": {
      "External id": 30,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 29
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821088.508, "dur": 0.471,
    "args": {
      "External id": 31,"Record function id": 0, "finished": true, "Ev Idx": 30
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821091.300, "dur": 1.168,
    "args": {
      "External id": 32,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 31
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821091.807, "dur": 0.232,
    "args": {
      "External id": 33,"Record function id": 0, "finished": true, "Ev Idx": 32
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564821101.706, "dur": 2.734,
    "args": {
      "External id": 34,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 33
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821103.508, "dur": 0.290,
    "args": {
      "External id": 35,"Record function id": 0, "finished": true, "Ev Idx": 34
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564821113.077, "dur": 21.590,
    "args": {
      "External id": 36,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 35
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 537, "tid": 537,
    "ts": 7429564821115.060, "dur": 3.180,
    "args": {
      "External id": 37,"Record function id": 0, "finished": true, "Ev Idx": 36
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564821119.387, "dur": 14.964,
    "args": {
      "External id": 38,"Record function id": 0, "finished": true, "Ev Idx": 37
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821138.969, "dur": 1.822,
    "args": {
      "External id": 39,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 38
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821139.524, "dur": 0.646,
    "args": {
      "External id": 40,"Record function id": 0, "finished": true, "Ev Idx": 39
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821141.448, "dur": 1.117,
    "args": {
      "External id": 41,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 40
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821141.917, "dur": 0.259,
    "args": {
      "External id": 42,"Record function id": 0, "finished": true, "Ev Idx": 41
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821143.201, "dur": 0.960,
    "args": {
      "External id": 43,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 42
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821143.607, "dur": 0.218,
    "args": {
      "External id": 44,"Record function id": 0, "finished": true, "Ev Idx": 43
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821146.409, "dur": 1.437,
    "args": {
      "External id": 45,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 44
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821146.931, "dur": 0.404,
    "args": {
      "External id": 46,"Record function id": 0, "finished": true, "Ev Idx": 45
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564821148.354, "dur": 1.576,
    "args": {
      "External id": 47,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 46
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821148.930, "dur": 0.365,
    "args": {
      "External id": 48,"Record function id": 0, "finished": true, "Ev Idx": 47
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564821150.373, "dur": 1.172,
    "args": {
      "External id": 49,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 48
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821150.815, "dur": 0.397,
    "args": {
      "External id": 50,"Record function id": 0, "finished": true, "Ev Idx": 49
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821152.097, "dur": 1.094,
    "args": {
      "External id": 51,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 50
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821152.549, "dur": 0.265,
    "args": {
      "External id": 52,"Record function id": 0, "finished": true, "Ev Idx": 51
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821158.399, "dur": 0.510,
    "args": {
      "External id": 53,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 52
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564821161.436, "dur": 13.896,
    "args": {
      "External id": 54,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 53
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 537, "tid": 537,
    "ts": 7429564821178.974, "dur": 15.092,
    "args": {
      "External id": 55,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 54
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821196.574, "dur": 1.675,
    "args": {
      "External id": 56,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 55
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821197.371, "dur": 0.411,
    "args": {
      "External id": 57,"Record function id": 0, "finished": true, "Ev Idx": 56
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821198.809, "dur": 1.096,
    "args": {
      "External id": 58,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 57
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821199.259, "dur": 0.244,
    "args": {
      "External id": 59,"Record function id": 0, "finished": true, "Ev Idx": 58
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821200.578, "dur": 0.940,
    "args": {
      "External id": 60,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 59
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821200.993, "dur": 0.194,
    "args": {
      "External id": 61,"Record function id": 0, "finished": true, "Ev Idx": 60
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill", "pid": 537, "tid": 537,
    "ts": 7429564821206.054, "dur": 32.992,
    "args": {
      "External id": 62,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 61
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564821208.320, "dur": 16.689,
    "args": {
      "External id": 63,"Record function id": 0, "finished": true, "Ev Idx": 62
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564821214.082, "dur": 2.843,
    "args": {
      "External id": 64,"Record function id": 0, "finished": true, "Ev Idx": 63
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564821215.433, "dur": 1.231,
    "args": {
      "External id": 65,"Record function id": 0, "finished": true, "Ev Idx": 64
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564821217.422, "dur": 7.436,
    "args": {
      "External id": 66,"Record function id": 0, "finished": true, "Ev Idx": 65
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 537, "tid": 537,
    "ts": 7429564821230.467, "dur": 8.208,
    "args": {
      "External id": 67,"Record function id": 0, "finished": true, "Ev Idx": 66
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821241.955, "dur": 1.402,
    "args": {
      "External id": 68,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 67
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821242.538, "dur": 0.403,
    "args": {
      "External id": 69,"Record function id": 0, "finished": true, "Ev Idx": 68
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821243.975, "dur": 2.433,
    "args": {
      "External id": 70,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 69
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821244.568, "dur": 1.457,
    "args": {
      "External id": 71,"Record function id": 0, "finished": true, "Ev Idx": 70
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821247.463, "dur": 1.733,
    "args": {
      "External id": 72,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 71
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821247.969, "dur": 0.851,
    "args": {
      "External id": 73,"Record function id": 0, "finished": true, "Ev Idx": 72
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564821250.245, "dur": 5.778,
    "args": {
      "External id": 74,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 73
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564821303.849, "dur": 1.955,
    "args": {
      "External id": 75,"Record function id": 0, "finished": true, "Ev Idx": 74
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821304.906, "dur": 0.433,
    "args": {
      "External id": 76,"Record function id": 0, "finished": true, "Ev Idx": 75
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821306.381, "dur": 1.532,
    "args": {
      "External id": 77,"Record function id": 0, "finished": true, "Ev Idx": 76
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821306.947, "dur": 0.411,
    "args": {
      "External id": 78,"Record function id": 0, "finished": true, "Ev Idx": 77
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564821308.498, "dur": 1.056,
    "args": {
      "External id": 79,"Record function id": 0, "finished": true, "Ev Idx": 78
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821308.993, "dur": 0.256,
    "args": {
      "External id": 80,"Record function id": 0, "finished": true, "Ev Idx": 79
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821312.830, "dur": 0.366,
    "args": {
      "External id": 81,"Record function id": 0, "finished": true, "Ev Idx": 80
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564821315.600, "dur": 2.082,
    "args": {
      "External id": 82,"Record function id": 0, "finished": true, "Ev Idx": 81
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821317.006, "dur": 0.266,
    "args": {
      "External id": 83,"Record function id": 0, "finished": true, "Ev Idx": 82
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821320.170, "dur": 0.309,
    "args": {
      "External id": 84,"Record function id": 0, "finished": true, "Ev Idx": 83
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821322.053, "dur": 1.539,
    "args": {
      "External id": 85,"Record function id": 0, "finished": true, "Ev Idx": 84
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821322.868, "dur": 0.272,
    "args": {
      "External id": 86,"Record function id": 0, "finished": true, "Ev Idx": 85
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564821324.065, "dur": 0.998,
    "args": {
      "External id": 87,"Record function id": 0, "finished": true, "Ev Idx": 86
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821324.539, "dur": 0.212,
    "args": {
      "External id": 88,"Record function id": 0, "finished": true, "Ev Idx": 87
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564821325.754, "dur": 1.247,
    "args": {
      "External id": 89,"Record function id": 0, "finished": true, "Ev Idx": 88
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821326.394, "dur": 0.249,
    "args": {
      "External id": 90,"Record function id": 0, "finished": true, "Ev Idx": 89
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821328.588, "dur": 14.071,
    "args": {
      "External id": 91,"Record function id": 0, "finished": true, "Ev Idx": 90
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 537, "tid": 537,
    "ts": 7429564821329.723, "dur": 12.788,
    "args": {
      "External id": 92,"Record function id": 0, "finished": true, "Ev Idx": 91
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 537, "tid": 537,
    "ts": 7429564821331.631, "dur": 1.764,
    "args": {
      "External id": 93,"Record function id": 0, "finished": true, "Ev Idx": 92
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564821334.053, "dur": 8.116,
    "args": {
      "External id": 94,"Record function id": 0, "finished": true, "Ev Idx": 93
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821399.690, "dur": 0.119,
    "args": {
      "External id": 95,"Record function id": 0, "finished": true, "Ev Idx": 94
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821400.950, "dur": 0.225,
    "args": {
      "External id": 96,"Record function id": 0, "finished": true, "Ev Idx": 95
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564821403.055, "dur": 76.476,
    "args": {
      "External id": 97,"Record function id": 0, "finished": true, "Ev Idx": 96
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564821404.382, "dur": 1.965,
    "args": {
      "External id": 98,"Record function id": 0, "finished": true, "Ev Idx": 97
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821405.607, "dur": 0.357,
    "args": {
      "External id": 99,"Record function id": 0, "finished": true, "Ev Idx": 98
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564821406.809, "dur": 1.563,
    "args": {
      "External id": 100,"Record function id": 0, "finished": true, "Ev Idx": 99
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564821407.156, "dur": 1.101,
    "args": {
      "External id": 101,"Record function id": 0, "finished": true, "Ev Idx": 100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564821408.918, "dur": 1.163,
    "args": {
      "External id": 102,"Record function id": 0, "finished": true, "Ev Idx": 101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821409.563, "dur": 0.216,
    "args": {
      "External id": 103,"Record function id": 0, "finished": true, "Ev Idx": 102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564821410.369, "dur": 0.958,
    "args": {
      "External id": 104,"Record function id": 0, "finished": true, "Ev Idx": 103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564821410.621, "dur": 0.644,
    "args": {
      "External id": 105,"Record function id": 0, "finished": true, "Ev Idx": 104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564821412.165, "dur": 64.085,
    "args": {
      "External id": 106,"Record function id": 0, "finished": true, "Ev Idx": 105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564821436.347, "dur": 1.985,
    "args": {
      "External id": 107,"Record function id": 0, "finished": true, "Ev Idx": 106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564821477.466, "dur": 1.460,
    "args": {
      "External id": 108,"Record function id": 0, "finished": true, "Ev Idx": 107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564821483.656, "dur": 2.884,
    "args": {
      "External id": 109,"Record function id": 0, "finished": true, "Ev Idx": 108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821485.192, "dur": 0.720,
    "args": {
      "External id": 110,"Record function id": 0, "finished": true, "Ev Idx": 109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564821492.909, "dur": 16.221,
    "args": {
      "External id": 111,"Record function id": 0, "finished": true, "Ev Idx": 110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cos", "pid": 537, "tid": 537,
    "ts": 7429564821514.062, "dur": 12.034,
    "args": {
      "External id": 112,"Record function id": 0, "finished": true, "Ev Idx": 111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564821530.830, "dur": 12.046,
    "args": {
      "External id": 113,"Record function id": 0, "finished": true, "Ev Idx": 112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sin", "pid": 537, "tid": 537,
    "ts": 7429564821547.217, "dur": 10.490,
    "args": {
      "External id": 114,"Record function id": 0, "finished": true, "Ev Idx": 113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564821559.623, "dur": 6.192,
    "args": {
      "External id": 115,"Record function id": 0, "finished": true, "Ev Idx": 114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821581.242, "dur": 0.142,
    "args": {
      "External id": 116,"Record function id": 0, "finished": true, "Ev Idx": 115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821583.265, "dur": 0.082,
    "args": {
      "External id": 117,"Record function id": 0, "finished": true, "Ev Idx": 116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821762.996, "dur": 0.423,
    "args": {
      "External id": 118,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564821767.234, "dur": 22.961,
    "args": {
      "External id": 119,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 292, "finished": true, "Ev Idx": 118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564821774.087, "dur": 0.421,
    "args": {
      "External id": 120,"Record function id": 0, "finished": true, "Ev Idx": 119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821774.925, "dur": 0.246,
    "args": {
      "External id": 121,"Record function id": 0, "finished": true, "Ev Idx": 120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564821800.466, "dur": 17.418,
    "args": {
      "External id": 122,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 293, "finished": true, "Ev Idx": 121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564821823.380, "dur": 13.185,
    "args": {
      "External id": 123,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 294, "finished": true, "Ev Idx": 122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564821840.110, "dur": 13.061,
    "args": {
      "External id": 124,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 295, "finished": true, "Ev Idx": 123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564821855.547, "dur": 9.828,
    "args": {
      "External id": 125,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 296, "finished": true, "Ev Idx": 124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564821869.024, "dur": 0.423,
    "args": {
      "External id": 126,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 297, "finished": true, "Ev Idx": 125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564821870.576, "dur": 8.392,
    "args": {
      "External id": 127,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 297, "finished": true, "Ev Idx": 126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564821910.052, "dur": 82.383,
    "args": {
      "External id": 128,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 298, "finished": true, "Ev Idx": 127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564821911.437, "dur": 10.863,
    "args": {
      "External id": 129,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 298, "finished": true, "Ev Idx": 128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564821913.828, "dur": 2.033,
    "args": {
      "External id": 130,"Record function id": 0, "finished": true, "Ev Idx": 129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564821914.899, "dur": 0.714,
    "args": {
      "External id": 131,"Record function id": 0, "finished": true, "Ev Idx": 130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564821923.032, "dur": 69.185,
    "args": {
      "External id": 132,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 299, "finished": true, "Ev Idx": 131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564821924.289, "dur": 3.219,
    "args": {
      "External id": 133,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 299, "finished": true, "Ev Idx": 132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564821924.799, "dur": 2.571,
    "args": {
      "External id": 134,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 299, "finished": true, "Ev Idx": 133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564821928.476, "dur": 60.045,
    "args": {
      "External id": 135,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 300, "finished": true, "Ev Idx": 134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564821955.024, "dur": 2.485,
    "args": {
      "External id": 136,"Record function id": 0, "finished": true, "Ev Idx": 135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564821989.431, "dur": 2.134,
    "args": {
      "External id": 137,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 301, "finished": true, "Ev Idx": 136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564821997.813, "dur": 1.580,
    "args": {
      "External id": 138,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 302, "finished": true, "Ev Idx": 137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564822001.647, "dur": 3.127,
    "args": {
      "External id": 139,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 303, "finished": true, "Ev Idx": 138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822003.026, "dur": 0.911,
    "args": {
      "External id": 140,"Record function id": 0, "finished": true, "Ev Idx": 139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564822015.014, "dur": 85.993,
    "args": {
      "External id": 141,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 304, "finished": true, "Ev Idx": 140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564822015.913, "dur": 3.046,
    "args": {
      "External id": 142,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 304, "finished": true, "Ev Idx": 141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564822017.207, "dur": 1.267,
    "args": {
      "External id": 143,"Record function id": 0, "finished": true, "Ev Idx": 142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822017.698, "dur": 0.667,
    "args": {
      "External id": 144,"Record function id": 0, "finished": true, "Ev Idx": 143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564822019.295, "dur": 81.589,
    "args": {
      "External id": 145,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 305, "finished": true, "Ev Idx": 144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822019.803, "dur": 2.046,
    "args": {
      "External id": 146,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 305, "finished": true, "Ev Idx": 145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564822020.224, "dur": 1.522,
    "args": {
      "External id": 147,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 305, "finished": true, "Ev Idx": 146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564822022.225, "dur": 74.433,
    "args": {
      "External id": 148,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 306, "finished": true, "Ev Idx": 147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564822037.203, "dur": 1.173,
    "args": {
      "External id": 149,"Record function id": 0, "finished": true, "Ev Idx": 148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564822098.297, "dur": 2.005,
    "args": {
      "External id": 150,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 307, "finished": true, "Ev Idx": 149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564822105.302, "dur": 2.158,
    "args": {
      "External id": 151,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 308, "finished": true, "Ev Idx": 150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564822109.459, "dur": 3.093,
    "args": {
      "External id": 152,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 309, "finished": true, "Ev Idx": 151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822110.657, "dur": 1.057,
    "args": {
      "External id": 153,"Record function id": 0, "finished": true, "Ev Idx": 152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564822122.048, "dur": 48.537,
    "args": {
      "External id": 154,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 310, "finished": true, "Ev Idx": 153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564822122.496, "dur": 3.359,
    "args": {
      "External id": 155,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 310, "finished": true, "Ev Idx": 154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564822124.110, "dur": 1.236,
    "args": {
      "External id": 156,"Record function id": 0, "finished": true, "Ev Idx": 155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822124.824, "dur": 0.392,
    "args": {
      "External id": 157,"Record function id": 0, "finished": true, "Ev Idx": 156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564822126.222, "dur": 44.246,
    "args": {
      "External id": 158,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 311, "finished": true, "Ev Idx": 157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822126.869, "dur": 2.141,
    "args": {
      "External id": 159,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 311, "finished": true, "Ev Idx": 158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564822127.187, "dur": 1.756,
    "args": {
      "External id": 160,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 311, "finished": true, "Ev Idx": 159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564822129.364, "dur": 39.004,
    "args": {
      "External id": 161,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 312, "finished": true, "Ev Idx": 160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564822142.408, "dur": 1.252,
    "args": {
      "External id": 162,"Record function id": 0, "finished": true, "Ev Idx": 161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564822168.982, "dur": 1.199,
    "args": {
      "External id": 163,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 313, "finished": true, "Ev Idx": 162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564822173.280, "dur": 1.344,
    "args": {
      "External id": 164,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 314, "finished": true, "Ev Idx": 163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564822176.205, "dur": 2.039,
    "args": {
      "External id": 165,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 315, "finished": true, "Ev Idx": 164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822176.934, "dur": 0.363,
    "args": {
      "External id": 166,"Record function id": 0, "finished": true, "Ev Idx": 165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564822182.987, "dur": 2.036,
    "args": {
      "External id": 167,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 316, "finished": true, "Ev Idx": 166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822184.054, "dur": 0.441,
    "args": {
      "External id": 168,"Record function id": 0, "finished": true, "Ev Idx": 167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564822186.388, "dur": 1.103,
    "args": {
      "External id": 169,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 316, "finished": true, "Ev Idx": 168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822186.843, "dur": 0.313,
    "args": {
      "External id": 170,"Record function id": 0, "finished": true, "Ev Idx": 169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564822188.651, "dur": 12.520,
    "args": {
      "External id": 171,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 316, "finished": true, "Ev Idx": 170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822208.270, "dur": 5.036,
    "args": {
      "External id": 172,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 317, "finished": true, "Ev Idx": 171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822211.114, "dur": 0.602,
    "args": {
      "External id": 173,"Record function id": 0, "finished": true, "Ev Idx": 172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822215.579, "dur": 3.481,
    "args": {
      "External id": 174,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 318, "finished": true, "Ev Idx": 173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822217.024, "dur": 1.248,
    "args": {
      "External id": 175,"Record function id": 0, "finished": true, "Ev Idx": 174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564822221.158, "dur": 12.108,
    "args": {
      "External id": 176,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 319, "finished": true, "Ev Idx": 175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564822238.025, "dur": 18.075,
    "args": {
      "External id": 177,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 320, "finished": true, "Ev Idx": 176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564822259.136, "dur": 9.522,
    "args": {
      "External id": 178,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 321, "finished": true, "Ev Idx": 177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564822271.271, "dur": 10.972,
    "args": {
      "External id": 179,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 322, "finished": true, "Ev Idx": 178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564822285.065, "dur": 7.643,
    "args": {
      "External id": 180,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 323, "finished": true, "Ev Idx": 179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822295.695, "dur": 2.932,
    "args": {
      "External id": 181,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 324, "finished": true, "Ev Idx": 180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822297.337, "dur": 0.608,
    "args": {
      "External id": 182,"Record function id": 0, "finished": true, "Ev Idx": 181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822300.337, "dur": 2.858,
    "args": {
      "External id": 183,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 325, "finished": true, "Ev Idx": 182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822301.962, "dur": 0.690,
    "args": {
      "External id": 184,"Record function id": 0, "finished": true, "Ev Idx": 183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564822304.256, "dur": 6.341,
    "args": {
      "External id": 185,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 326, "finished": true, "Ev Idx": 184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564822312.620, "dur": 9.833,
    "args": {
      "External id": 186,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 327, "finished": true, "Ev Idx": 185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564822324.900, "dur": 7.489,
    "args": {
      "External id": 187,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 328, "finished": true, "Ev Idx": 186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564822333.773, "dur": 8.214,
    "args": {
      "External id": 188,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 329, "finished": true, "Ev Idx": 187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822377.970, "dur": 3.085,
    "args": {
      "External id": 189,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 330, "finished": true, "Ev Idx": 188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822379.685, "dur": 0.659,
    "args": {
      "External id": 190,"Record function id": 0, "finished": true, "Ev Idx": 189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822381.692, "dur": 2.564,
    "args": {
      "External id": 191,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 331, "finished": true, "Ev Idx": 190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822383.234, "dur": 0.522,
    "args": {
      "External id": 192,"Record function id": 0, "finished": true, "Ev Idx": 191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564822385.034, "dur": 3.159,
    "args": {
      "External id": 193,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 332, "finished": true, "Ev Idx": 192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822386.482, "dur": 0.792,
    "args": {
      "External id": 194,"Record function id": 0, "finished": true, "Ev Idx": 193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822388.938, "dur": 2.593,
    "args": {
      "External id": 195,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 333, "finished": true, "Ev Idx": 194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822390.578, "dur": 0.431,
    "args": {
      "External id": 196,"Record function id": 0, "finished": true, "Ev Idx": 195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822392.178, "dur": 1.635,
    "args": {
      "External id": 197,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 334, "finished": true, "Ev Idx": 196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822393.052, "dur": 0.313,
    "args": {
      "External id": 198,"Record function id": 0, "finished": true, "Ev Idx": 197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564822397.176, "dur": 3.176,
    "args": {
      "External id": 199,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 335, "finished": true, "Ev Idx": 198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822399.129, "dur": 0.421,
    "args": {
      "External id": 200,"Record function id": 0, "finished": true, "Ev Idx": 199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822403.187, "dur": 23.482,
    "args": {
      "External id": 201,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 336, "finished": true, "Ev Idx": 200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564822406.188, "dur": 17.879,
    "args": {
      "External id": 202,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 336, "finished": true, "Ev Idx": 201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564822408.406, "dur": 2.943,
    "args": {
      "External id": 203,"Record function id": 0, "finished": true, "Ev Idx": 202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564822409.203, "dur": 1.893,
    "args": {
      "External id": 204,"Record function id": 0, "finished": true, "Ev Idx": 203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564822412.228, "dur": 11.037,
    "args": {
      "External id": 205,"Record function id": 0, "finished": true, "Ev Idx": 204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564822424.631, "dur": 1.793,
    "args": {
      "External id": 206,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 337, "finished": true, "Ev Idx": 205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822430.296, "dur": 2.417,
    "args": {
      "External id": 207,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 338, "finished": true, "Ev Idx": 206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822431.654, "dur": 0.487,
    "args": {
      "External id": 208,"Record function id": 0, "finished": true, "Ev Idx": 207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822433.295, "dur": 1.948,
    "args": {
      "External id": 209,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 339, "finished": true, "Ev Idx": 208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822433.998, "dur": 0.683,
    "args": {
      "External id": 210,"Record function id": 0, "finished": true, "Ev Idx": 209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564822435.841, "dur": 1.760,
    "args": {
      "External id": 211,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 340, "finished": true, "Ev Idx": 210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822436.536, "dur": 0.535,
    "args": {
      "External id": 212,"Record function id": 0, "finished": true, "Ev Idx": 211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822438.276, "dur": 1.871,
    "args": {
      "External id": 213,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 341, "finished": true, "Ev Idx": 212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822439.298, "dur": 0.379,
    "args": {
      "External id": 214,"Record function id": 0, "finished": true, "Ev Idx": 213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822440.734, "dur": 1.490,
    "args": {
      "External id": 215,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 342, "finished": true, "Ev Idx": 214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822441.490, "dur": 0.273,
    "args": {
      "External id": 216,"Record function id": 0, "finished": true, "Ev Idx": 215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564822443.816, "dur": 2.045,
    "args": {
      "External id": 217,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 343, "finished": true, "Ev Idx": 216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822445.017, "dur": 0.362,
    "args": {
      "External id": 218,"Record function id": 0, "finished": true, "Ev Idx": 217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822447.438, "dur": 13.219,
    "args": {
      "External id": 219,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 344, "finished": true, "Ev Idx": 218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564822448.521, "dur": 10.313,
    "args": {
      "External id": 220,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 344, "finished": true, "Ev Idx": 219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564822450.057, "dur": 1.829,
    "args": {
      "External id": 221,"Record function id": 0, "finished": true, "Ev Idx": 220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564822450.475, "dur": 1.302,
    "args": {
      "External id": 222,"Record function id": 0, "finished": true, "Ev Idx": 221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564822452.331, "dur": 6.144,
    "args": {
      "External id": 223,"Record function id": 0, "finished": true, "Ev Idx": 222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564822459.377, "dur": 1.008,
    "args": {
      "External id": 224,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 345, "finished": true, "Ev Idx": 223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564822463.003, "dur": 1.796,
    "args": {
      "External id": 225,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 346, "finished": true, "Ev Idx": 224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822463.855, "dur": 0.365,
    "args": {
      "External id": 226,"Record function id": 0, "finished": true, "Ev Idx": 225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564822467.862, "dur": 55.316,
    "args": {
      "External id": 227,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 347, "finished": true, "Ev Idx": 226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564822468.471, "dur": 1.713,
    "args": {
      "External id": 228,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 347, "finished": true, "Ev Idx": 227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822469.282, "dur": 0.455,
    "args": {
      "External id": 229,"Record function id": 0, "finished": true, "Ev Idx": 228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822470.571, "dur": 9.365,
    "args": {
      "External id": 230,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 348, "finished": true, "Ev Idx": 229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564822477.538, "dur": 2.247,
    "args": {
      "External id": 231,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 348, "finished": true, "Ev Idx": 230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564822480.511, "dur": 1.375,
    "args": {
      "External id": 232,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 349, "finished": true, "Ev Idx": 231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822481.160, "dur": 0.339,
    "args": {
      "External id": 233,"Record function id": 0, "finished": true, "Ev Idx": 232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822482.214, "dur": 2.544,
    "args": {
      "External id": 234,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 350, "finished": true, "Ev Idx": 233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564822483.040, "dur": 1.574,
    "args": {
      "External id": 235,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 350, "finished": true, "Ev Idx": 234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564822485.477, "dur": 35.343,
    "args": {
      "External id": 236,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 351, "finished": true, "Ev Idx": 235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564822501.645, "dur": 1.166,
    "args": {
      "External id": 237,"Record function id": 0, "finished": true, "Ev Idx": 236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564822521.557, "dur": 1.172,
    "args": {
      "External id": 238,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 352, "finished": true, "Ev Idx": 237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564822527.775, "dur": 10.451,
    "args": {
      "External id": 239,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 353, "finished": true, "Ev Idx": 238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822541.552, "dur": 1.909,
    "args": {
      "External id": 240,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 354, "finished": true, "Ev Idx": 239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822542.285, "dur": 0.529,
    "args": {
      "External id": 241,"Record function id": 0, "finished": true, "Ev Idx": 240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822544.096, "dur": 2.248,
    "args": {
      "External id": 242,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 354, "finished": true, "Ev Idx": 241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822544.676, "dur": 1.282,
    "args": {
      "External id": 243,"Record function id": 0, "finished": true, "Ev Idx": 242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564822547.032, "dur": 1.206,
    "args": {
      "External id": 244,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 354, "finished": true, "Ev Idx": 243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822547.516, "dur": 0.401,
    "args": {
      "External id": 245,"Record function id": 0, "finished": true, "Ev Idx": 244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564822549.971, "dur": 7.953,
    "args": {
      "External id": 246,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 354, "finished": true, "Ev Idx": 245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564822567.941, "dur": 16.163,
    "args": {
      "External id": 247,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 355, "finished": true, "Ev Idx": 246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564822569.337, "dur": 0.327,
    "args": {
      "External id": 248,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 355, "finished": true, "Ev Idx": 247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564822570.908, "dur": 12.922,
    "args": {
      "External id": 249,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 355, "finished": true, "Ev Idx": 248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564822587.375, "dur": 0.248,
    "args": {
      "External id": 250,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 356, "finished": true, "Ev Idx": 249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564822602.036, "dur": 0.516,
    "args": {
      "External id": 251,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 356, "finished": true, "Ev Idx": 250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564822604.648, "dur": 45.496,
    "args": {
      "External id": 252,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 356, "finished": true, "Ev Idx": 251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564822605.522, "dur": 2.275,
    "args": {
      "External id": 253,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 356, "finished": true, "Ev Idx": 252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822606.839, "dur": 0.439,
    "args": {
      "External id": 254,"Record function id": 0, "finished": true, "Ev Idx": 253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822608.177, "dur": 2.148,
    "args": {
      "External id": 255,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 357, "finished": true, "Ev Idx": 254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564822608.579, "dur": 1.621,
    "args": {
      "External id": 256,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 357, "finished": true, "Ev Idx": 255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564822610.765, "dur": 1.654,
    "args": {
      "External id": 257,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 358, "finished": true, "Ev Idx": 256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822611.719, "dur": 0.342,
    "args": {
      "External id": 258,"Record function id": 0, "finished": true, "Ev Idx": 257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822612.696, "dur": 2.169,
    "args": {
      "External id": 259,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 359, "finished": true, "Ev Idx": 258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564822612.971, "dur": 1.812,
    "args": {
      "External id": 260,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 359, "finished": true, "Ev Idx": 259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564822615.291, "dur": 32.881,
    "args": {
      "External id": 261,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 360, "finished": true, "Ev Idx": 260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564822628.310, "dur": 1.198,
    "args": {
      "External id": 262,"Record function id": 0, "finished": true, "Ev Idx": 261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564822648.771, "dur": 1.067,
    "args": {
      "External id": 263,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 361, "finished": true, "Ev Idx": 262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564822652.779, "dur": 1.924,
    "args": {
      "External id": 264,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 362, "finished": true, "Ev Idx": 263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822653.808, "dur": 0.370,
    "args": {
      "External id": 265,"Record function id": 0, "finished": true, "Ev Idx": 264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564822659.184, "dur": 11.736,
    "args": {
      "External id": 266,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 363, "finished": true, "Ev Idx": 265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564822659.758, "dur": 11.012,
    "args": {
      "External id": 267,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 363, "finished": true, "Ev Idx": 266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564822660.819, "dur": 2.120,
    "args": {
      "External id": 268,"Record function id": 0, "finished": true, "Ev Idx": 267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564822661.306, "dur": 1.495,
    "args": {
      "External id": 269,"Record function id": 0, "finished": true, "Ev Idx": 268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564822663.315, "dur": 7.008,
    "args": {
      "External id": 270,"Record function id": 0, "finished": true, "Ev Idx": 269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822677.737, "dur": 1.922,
    "args": {
      "External id": 271,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 364, "finished": true, "Ev Idx": 270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564822678.139, "dur": 1.406,
    "args": {
      "External id": 272,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 364, "finished": true, "Ev Idx": 271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564822693.605, "dur": 38.875,
    "args": {
      "External id": 273,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 365, "finished": true, "Ev Idx": 272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564822694.190, "dur": 3.860,
    "args": {
      "External id": 274,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 365, "finished": true, "Ev Idx": 273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564822696.015, "dur": 1.424,
    "args": {
      "External id": 275,"Record function id": 0, "finished": true, "Ev Idx": 274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822696.936, "dur": 0.368,
    "args": {
      "External id": 276,"Record function id": 0, "finished": true, "Ev Idx": 275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564822698.575, "dur": 33.728,
    "args": {
      "External id": 277,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 366, "finished": true, "Ev Idx": 276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822699.247, "dur": 1.624,
    "args": {
      "External id": 278,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 366, "finished": true, "Ev Idx": 277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564822699.504, "dur": 1.262,
    "args": {
      "External id": 279,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 366, "finished": true, "Ev Idx": 278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564822701.289, "dur": 28.704,
    "args": {
      "External id": 280,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 367, "finished": true, "Ev Idx": 279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564822713.799, "dur": 1.831,
    "args": {
      "External id": 281,"Record function id": 0, "finished": true, "Ev Idx": 280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564822730.632, "dur": 1.370,
    "args": {
      "External id": 282,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 368, "finished": true, "Ev Idx": 281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564822738.008, "dur": 8.948,
    "args": {
      "External id": 283,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 369, "finished": true, "Ev Idx": 282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564822757.170, "dur": 0.278,
    "args": {
      "External id": 284,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 370, "finished": true, "Ev Idx": 283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564822759.961, "dur": 14.361,
    "args": {
      "External id": 285,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 370, "finished": true, "Ev Idx": 284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564822763.037, "dur": 0.456,
    "args": {
      "External id": 286,"Record function id": 0, "finished": true, "Ev Idx": 285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564822764.108, "dur": 0.068,
    "args": {
      "External id": 287,"Record function id": 0, "finished": true, "Ev Idx": 286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564822777.636, "dur": 14.427,
    "args": {
      "External id": 288,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 371, "finished": true, "Ev Idx": 287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564822795.876, "dur": 8.177,
    "args": {
      "External id": 289,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 372, "finished": true, "Ev Idx": 288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564822806.501, "dur": 10.136,
    "args": {
      "External id": 290,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 373, "finished": true, "Ev Idx": 289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564822818.456, "dur": 9.979,
    "args": {
      "External id": 291,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 374, "finished": true, "Ev Idx": 290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564822831.419, "dur": 0.107,
    "args": {
      "External id": 292,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 375, "finished": true, "Ev Idx": 291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564822832.443, "dur": 8.066,
    "args": {
      "External id": 293,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 375, "finished": true, "Ev Idx": 292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564822859.810, "dur": 51.517,
    "args": {
      "External id": 294,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 376, "finished": true, "Ev Idx": 293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564822860.552, "dur": 4.169,
    "args": {
      "External id": 295,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 376, "finished": true, "Ev Idx": 294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564822862.618, "dur": 1.468,
    "args": {
      "External id": 296,"Record function id": 0, "finished": true, "Ev Idx": 295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822863.469, "dur": 0.469,
    "args": {
      "External id": 297,"Record function id": 0, "finished": true, "Ev Idx": 296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564822865.210, "dur": 46.011,
    "args": {
      "External id": 298,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 377, "finished": true, "Ev Idx": 297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822865.758, "dur": 1.575,
    "args": {
      "External id": 299,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 377, "finished": true, "Ev Idx": 298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564822866.120, "dur": 1.112,
    "args": {
      "External id": 300,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 377, "finished": true, "Ev Idx": 299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564822867.840, "dur": 40.879,
    "args": {
      "External id": 301,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 378, "finished": true, "Ev Idx": 300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564822879.372, "dur": 1.838,
    "args": {
      "External id": 302,"Record function id": 0, "finished": true, "Ev Idx": 301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564822909.550, "dur": 1.408,
    "args": {
      "External id": 303,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 379, "finished": true, "Ev Idx": 302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564822925.579, "dur": 12.361,
    "args": {
      "External id": 304,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 380, "finished": true, "Ev Idx": 303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564822945.598, "dur": 40.405,
    "args": {
      "External id": 305,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 381, "finished": true, "Ev Idx": 304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564822946.073, "dur": 3.311,
    "args": {
      "External id": 306,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 381, "finished": true, "Ev Idx": 305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564822947.238, "dur": 1.586,
    "args": {
      "External id": 307,"Record function id": 0, "finished": true, "Ev Idx": 306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564822947.829, "dur": 0.869,
    "args": {
      "External id": 308,"Record function id": 0, "finished": true, "Ev Idx": 307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564822949.717, "dur": 36.208,
    "args": {
      "External id": 309,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 382, "finished": true, "Ev Idx": 308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564822950.243, "dur": 1.829,
    "args": {
      "External id": 310,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 382, "finished": true, "Ev Idx": 309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564822950.700, "dur": 1.261,
    "args": {
      "External id": 311,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 382, "finished": true, "Ev Idx": 310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564822952.417, "dur": 31.620,
    "args": {
      "External id": 312,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 383, "finished": true, "Ev Idx": 311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564822962.817, "dur": 1.114,
    "args": {
      "External id": 313,"Record function id": 0, "finished": true, "Ev Idx": 312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564822984.628, "dur": 0.989,
    "args": {
      "External id": 314,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 384, "finished": true, "Ev Idx": 313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564822987.832, "dur": 9.085,
    "args": {
      "External id": 315,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 385, "finished": true, "Ev Idx": 314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564823005.681, "dur": 45.787,
    "args": {
      "External id": 316,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 386, "finished": true, "Ev Idx": 315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564823006.011, "dur": 2.613,
    "args": {
      "External id": 317,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 386, "finished": true, "Ev Idx": 316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564823007.226, "dur": 0.864,
    "args": {
      "External id": 318,"Record function id": 0, "finished": true, "Ev Idx": 317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823007.606, "dur": 0.359,
    "args": {
      "External id": 319,"Record function id": 0, "finished": true, "Ev Idx": 318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564823008.911, "dur": 42.435,
    "args": {
      "External id": 320,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 387, "finished": true, "Ev Idx": 319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823009.498, "dur": 1.786,
    "args": {
      "External id": 321,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 387, "finished": true, "Ev Idx": 320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823009.775, "dur": 1.420,
    "args": {
      "External id": 322,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 387, "finished": true, "Ev Idx": 321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564823011.603, "dur": 37.806,
    "args": {
      "External id": 323,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 388, "finished": true, "Ev Idx": 322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564823021.365, "dur": 1.292,
    "args": {
      "External id": 324,"Record function id": 0, "finished": true, "Ev Idx": 323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564823049.984, "dur": 1.096,
    "args": {
      "External id": 325,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 389, "finished": true, "Ev Idx": 324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564823054.356, "dur": 8.108,
    "args": {
      "External id": 326,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 390, "finished": true, "Ev Idx": 325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564823082.598, "dur": 0.145,
    "args": {
      "External id": 327,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 391, "finished": true, "Ev Idx": 326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564823084.627, "dur": 10.454,
    "args": {
      "External id": 328,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 391, "finished": true, "Ev Idx": 327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564823086.565, "dur": 0.343,
    "args": {
      "External id": 329,"Record function id": 0, "finished": true, "Ev Idx": 328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564823087.309, "dur": 0.071,
    "args": {
      "External id": 330,"Record function id": 0, "finished": true, "Ev Idx": 329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564823097.589, "dur": 12.048,
    "args": {
      "External id": 331,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 392, "finished": true, "Ev Idx": 330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564823112.832, "dur": 8.266,
    "args": {
      "External id": 332,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 393, "finished": true, "Ev Idx": 331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564823122.750, "dur": 8.560,
    "args": {
      "External id": 333,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 394, "finished": true, "Ev Idx": 332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564823132.872, "dur": 10.109,
    "args": {
      "External id": 334,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 395, "finished": true, "Ev Idx": 333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564823145.262, "dur": 0.101,
    "args": {
      "External id": 335,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 396, "finished": true, "Ev Idx": 334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564823146.199, "dur": 8.865,
    "args": {
      "External id": 336,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 396, "finished": true, "Ev Idx": 335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564823172.360, "dur": 38.550,
    "args": {
      "External id": 337,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 397, "finished": true, "Ev Idx": 336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564823172.787, "dur": 3.157,
    "args": {
      "External id": 338,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 397, "finished": true, "Ev Idx": 337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564823174.156, "dur": 1.142,
    "args": {
      "External id": 339,"Record function id": 0, "finished": true, "Ev Idx": 338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823174.648, "dur": 0.522,
    "args": {
      "External id": 340,"Record function id": 0, "finished": true, "Ev Idx": 339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564823176.226, "dur": 34.584,
    "args": {
      "External id": 341,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 398, "finished": true, "Ev Idx": 340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823176.664, "dur": 2.223,
    "args": {
      "External id": 342,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 398, "finished": true, "Ev Idx": 341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823176.996, "dur": 1.805,
    "args": {
      "External id": 343,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 398, "finished": true, "Ev Idx": 342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564823179.214, "dur": 29.444,
    "args": {
      "External id": 344,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 399, "finished": true, "Ev Idx": 343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564823191.352, "dur": 1.243,
    "args": {
      "External id": 345,"Record function id": 0, "finished": true, "Ev Idx": 344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564823209.437, "dur": 1.048,
    "args": {
      "External id": 346,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 400, "finished": true, "Ev Idx": 345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823213.919, "dur": 1.381,
    "args": {
      "External id": 347,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 401, "finished": true, "Ev Idx": 346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564823217.206, "dur": 2.419,
    "args": {
      "External id": 348,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 402, "finished": true, "Ev Idx": 347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823218.298, "dur": 0.666,
    "args": {
      "External id": 349,"Record function id": 0, "finished": true, "Ev Idx": 348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564823227.096, "dur": 40.579,
    "args": {
      "External id": 350,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 403, "finished": true, "Ev Idx": 349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564823227.435, "dur": 2.379,
    "args": {
      "External id": 351,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 403, "finished": true, "Ev Idx": 350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564823228.411, "dur": 0.965,
    "args": {
      "External id": 352,"Record function id": 0, "finished": true, "Ev Idx": 351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823228.992, "dur": 0.281,
    "args": {
      "External id": 353,"Record function id": 0, "finished": true, "Ev Idx": 352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564823230.135, "dur": 37.423,
    "args": {
      "External id": 354,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 404, "finished": true, "Ev Idx": 353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823230.754, "dur": 1.698,
    "args": {
      "External id": 355,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 404, "finished": true, "Ev Idx": 354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823231.379, "dur": 0.969,
    "args": {
      "External id": 356,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 404, "finished": true, "Ev Idx": 355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564823232.765, "dur": 32.957,
    "args": {
      "External id": 357,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 405, "finished": true, "Ev Idx": 356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564823242.267, "dur": 1.297,
    "args": {
      "External id": 358,"Record function id": 0, "finished": true, "Ev Idx": 357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564823266.301, "dur": 1.002,
    "args": {
      "External id": 359,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 406, "finished": true, "Ev Idx": 358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823269.803, "dur": 1.430,
    "args": {
      "External id": 360,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 407, "finished": true, "Ev Idx": 359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564823272.608, "dur": 2.374,
    "args": {
      "External id": 361,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 408, "finished": true, "Ev Idx": 360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823273.379, "dur": 1.047,
    "args": {
      "External id": 362,"Record function id": 0, "finished": true, "Ev Idx": 361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564823281.583, "dur": 38.722,
    "args": {
      "External id": 363,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 409, "finished": true, "Ev Idx": 362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564823282.045, "dur": 2.207,
    "args": {
      "External id": 364,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 409, "finished": true, "Ev Idx": 363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564823282.822, "dur": 0.950,
    "args": {
      "External id": 365,"Record function id": 0, "finished": true, "Ev Idx": 364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823283.422, "dur": 0.257,
    "args": {
      "External id": 366,"Record function id": 0, "finished": true, "Ev Idx": 365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564823284.526, "dur": 35.687,
    "args": {
      "External id": 367,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 410, "finished": true, "Ev Idx": 366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823284.906, "dur": 1.921,
    "args": {
      "External id": 368,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 410, "finished": true, "Ev Idx": 367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823285.671, "dur": 1.094,
    "args": {
      "External id": 369,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 410, "finished": true, "Ev Idx": 368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564823287.215, "dur": 31.218,
    "args": {
      "External id": 370,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 411, "finished": true, "Ev Idx": 369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564823297.347, "dur": 1.136,
    "args": {
      "External id": 371,"Record function id": 0, "finished": true, "Ev Idx": 370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564823319.062, "dur": 0.874,
    "args": {
      "External id": 372,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 412, "finished": true, "Ev Idx": 371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823322.594, "dur": 1.349,
    "args": {
      "External id": 373,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 413, "finished": true, "Ev Idx": 372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564823325.538, "dur": 1.762,
    "args": {
      "External id": 374,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 414, "finished": true, "Ev Idx": 373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823326.225, "dur": 0.577,
    "args": {
      "External id": 375,"Record function id": 0, "finished": true, "Ev Idx": 374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564823329.913, "dur": 2.188,
    "args": {
      "External id": 376,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 415, "finished": true, "Ev Idx": 375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823331.186, "dur": 0.326,
    "args": {
      "External id": 377,"Record function id": 0, "finished": true, "Ev Idx": 376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564823333.183, "dur": 1.196,
    "args": {
      "External id": 378,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 415, "finished": true, "Ev Idx": 377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823333.646, "dur": 0.414,
    "args": {
      "External id": 379,"Record function id": 0, "finished": true, "Ev Idx": 378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564823335.247, "dur": 10.333,
    "args": {
      "External id": 380,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 415, "finished": true, "Ev Idx": 379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823350.361, "dur": 3.482,
    "args": {
      "External id": 381,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 416, "finished": true, "Ev Idx": 380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823352.149, "dur": 0.538,
    "args": {
      "External id": 382,"Record function id": 0, "finished": true, "Ev Idx": 381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823356.041, "dur": 2.981,
    "args": {
      "External id": 383,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 417, "finished": true, "Ev Idx": 382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823357.348, "dur": 0.538,
    "args": {
      "External id": 384,"Record function id": 0, "finished": true, "Ev Idx": 383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564823360.774, "dur": 9.545,
    "args": {
      "External id": 385,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 418, "finished": true, "Ev Idx": 384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564823373.800, "dur": 14.830,
    "args": {
      "External id": 386,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 419, "finished": true, "Ev Idx": 385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564823391.056, "dur": 9.658,
    "args": {
      "External id": 387,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 420, "finished": true, "Ev Idx": 386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564823402.602, "dur": 8.407,
    "args": {
      "External id": 388,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 421, "finished": true, "Ev Idx": 387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564823413.241, "dur": 8.436,
    "args": {
      "External id": 389,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 422, "finished": true, "Ev Idx": 388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823423.873, "dur": 2.627,
    "args": {
      "External id": 390,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 423, "finished": true, "Ev Idx": 389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823425.398, "dur": 0.446,
    "args": {
      "External id": 391,"Record function id": 0, "finished": true, "Ev Idx": 390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823428.035, "dur": 2.351,
    "args": {
      "External id": 392,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 424, "finished": true, "Ev Idx": 391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823429.574, "dur": 0.318,
    "args": {
      "External id": 393,"Record function id": 0, "finished": true, "Ev Idx": 392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564823431.411, "dur": 7.831,
    "args": {
      "External id": 394,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 425, "finished": true, "Ev Idx": 393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564823440.904, "dur": 10.957,
    "args": {
      "External id": 395,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 426, "finished": true, "Ev Idx": 394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564823454.180, "dur": 7.951,
    "args": {
      "External id": 396,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 427, "finished": true, "Ev Idx": 395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564823463.398, "dur": 7.238,
    "args": {
      "External id": 397,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 428, "finished": true, "Ev Idx": 396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823492.637, "dur": 3.543,
    "args": {
      "External id": 398,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 429, "finished": true, "Ev Idx": 397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823494.938, "dur": 0.521,
    "args": {
      "External id": 399,"Record function id": 0, "finished": true, "Ev Idx": 398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823496.798, "dur": 1.699,
    "args": {
      "External id": 400,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 430, "finished": true, "Ev Idx": 399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823497.666, "dur": 0.353,
    "args": {
      "External id": 401,"Record function id": 0, "finished": true, "Ev Idx": 400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564823499.152, "dur": 2.140,
    "args": {
      "External id": 402,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 431, "finished": true, "Ev Idx": 401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823500.196, "dur": 0.512,
    "args": {
      "External id": 403,"Record function id": 0, "finished": true, "Ev Idx": 402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823501.956, "dur": 1.903,
    "args": {
      "External id": 404,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 432, "finished": true, "Ev Idx": 403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823502.737, "dur": 0.591,
    "args": {
      "External id": 405,"Record function id": 0, "finished": true, "Ev Idx": 404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823504.457, "dur": 1.830,
    "args": {
      "External id": 406,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 433, "finished": true, "Ev Idx": 405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823505.365, "dur": 0.446,
    "args": {
      "External id": 407,"Record function id": 0, "finished": true, "Ev Idx": 406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564823509.213, "dur": 2.839,
    "args": {
      "External id": 408,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 434, "finished": true, "Ev Idx": 407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823510.953, "dur": 0.447,
    "args": {
      "External id": 409,"Record function id": 0, "finished": true, "Ev Idx": 408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823514.552, "dur": 18.966,
    "args": {
      "External id": 410,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 435, "finished": true, "Ev Idx": 409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564823516.755, "dur": 14.648,
    "args": {
      "External id": 411,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 435, "finished": true, "Ev Idx": 410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564823518.942, "dur": 2.668,
    "args": {
      "External id": 412,"Record function id": 0, "finished": true, "Ev Idx": 411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564823519.696, "dur": 1.701,
    "args": {
      "External id": 413,"Record function id": 0, "finished": true, "Ev Idx": 412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564823522.079, "dur": 8.708,
    "args": {
      "External id": 414,"Record function id": 0, "finished": true, "Ev Idx": 413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564823531.968, "dur": 1.320,
    "args": {
      "External id": 415,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 436, "finished": true, "Ev Idx": 414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823536.664, "dur": 1.894,
    "args": {
      "External id": 416,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 437, "finished": true, "Ev Idx": 415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823537.571, "dur": 0.430,
    "args": {
      "External id": 417,"Record function id": 0, "finished": true, "Ev Idx": 416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823539.172, "dur": 1.440,
    "args": {
      "External id": 418,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 438, "finished": true, "Ev Idx": 417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823539.895, "dur": 0.287,
    "args": {
      "External id": 419,"Record function id": 0, "finished": true, "Ev Idx": 418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564823541.152, "dur": 1.457,
    "args": {
      "External id": 420,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 439, "finished": true, "Ev Idx": 419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823541.851, "dur": 0.316,
    "args": {
      "External id": 421,"Record function id": 0, "finished": true, "Ev Idx": 420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823543.218, "dur": 2.435,
    "args": {
      "External id": 422,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 440, "finished": true, "Ev Idx": 421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823544.753, "dur": 0.392,
    "args": {
      "External id": 423,"Record function id": 0, "finished": true, "Ev Idx": 422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823546.242, "dur": 2.108,
    "args": {
      "External id": 424,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 441, "finished": true, "Ev Idx": 423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823547.473, "dur": 0.425,
    "args": {
      "External id": 425,"Record function id": 0, "finished": true, "Ev Idx": 424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564823549.985, "dur": 1.643,
    "args": {
      "External id": 426,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 442, "finished": true, "Ev Idx": 425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823550.894, "dur": 0.327,
    "args": {
      "External id": 427,"Record function id": 0, "finished": true, "Ev Idx": 426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823552.970, "dur": 14.061,
    "args": {
      "External id": 428,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 443, "finished": true, "Ev Idx": 427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564823554.079, "dur": 11.277,
    "args": {
      "External id": 429,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 443, "finished": true, "Ev Idx": 428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564823554.720, "dur": 1.779,
    "args": {
      "External id": 430,"Record function id": 0, "finished": true, "Ev Idx": 429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564823555.054, "dur": 1.319,
    "args": {
      "External id": 431,"Record function id": 0, "finished": true, "Ev Idx": 430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564823556.843, "dur": 8.109,
    "args": {
      "External id": 432,"Record function id": 0, "finished": true, "Ev Idx": 431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564823565.791, "dur": 1.087,
    "args": {
      "External id": 433,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 444, "finished": true, "Ev Idx": 432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564823568.977, "dur": 2.220,
    "args": {
      "External id": 434,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 445, "finished": true, "Ev Idx": 433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823570.280, "dur": 0.419,
    "args": {
      "External id": 435,"Record function id": 0, "finished": true, "Ev Idx": 434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564823572.737, "dur": 45.259,
    "args": {
      "External id": 436,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 446, "finished": true, "Ev Idx": 435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564823573.602, "dur": 1.807,
    "args": {
      "External id": 437,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 446, "finished": true, "Ev Idx": 436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823574.665, "dur": 0.322,
    "args": {
      "External id": 438,"Record function id": 0, "finished": true, "Ev Idx": 437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823575.899, "dur": 3.580,
    "args": {
      "External id": 439,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 447, "finished": true, "Ev Idx": 438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564823577.682, "dur": 1.670,
    "args": {
      "External id": 440,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 447, "finished": true, "Ev Idx": 439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564823579.958, "dur": 1.344,
    "args": {
      "External id": 441,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 448, "finished": true, "Ev Idx": 440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823580.636, "dur": 0.287,
    "args": {
      "External id": 442,"Record function id": 0, "finished": true, "Ev Idx": 441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823581.574, "dur": 1.815,
    "args": {
      "External id": 443,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 449, "finished": true, "Ev Idx": 442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564823582.407, "dur": 0.881,
    "args": {
      "External id": 444,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 449, "finished": true, "Ev Idx": 443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564823583.992, "dur": 31.950,
    "args": {
      "External id": 445,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 450, "finished": true, "Ev Idx": 444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564823595.779, "dur": 1.073,
    "args": {
      "External id": 446,"Record function id": 0, "finished": true, "Ev Idx": 445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564823616.516, "dur": 1.168,
    "args": {
      "External id": 447,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 451, "finished": true, "Ev Idx": 446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564823620.970, "dur": 10.487,
    "args": {
      "External id": 448,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 452, "finished": true, "Ev Idx": 447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823634.638, "dur": 1.868,
    "args": {
      "External id": 449,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 453, "finished": true, "Ev Idx": 448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823635.450, "dur": 0.417,
    "args": {
      "External id": 450,"Record function id": 0, "finished": true, "Ev Idx": 449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823637.148, "dur": 1.227,
    "args": {
      "External id": 451,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 453, "finished": true, "Ev Idx": 450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823637.703, "dur": 0.302,
    "args": {
      "External id": 452,"Record function id": 0, "finished": true, "Ev Idx": 451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564823639.007, "dur": 1.187,
    "args": {
      "External id": 453,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 453, "finished": true, "Ev Idx": 452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823639.487, "dur": 0.393,
    "args": {
      "External id": 454,"Record function id": 0, "finished": true, "Ev Idx": 453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564823641.755, "dur": 8.255,
    "args": {
      "External id": 455,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 453, "finished": true, "Ev Idx": 454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564823655.026, "dur": 13.270,
    "args": {
      "External id": 456,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 454, "finished": true, "Ev Idx": 455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564823656.018, "dur": 0.281,
    "args": {
      "External id": 457,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 454, "finished": true, "Ev Idx": 456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564823657.095, "dur": 10.934,
    "args": {
      "External id": 458,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 454, "finished": true, "Ev Idx": 457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564823670.913, "dur": 0.200,
    "args": {
      "External id": 459,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 455, "finished": true, "Ev Idx": 458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564823682.129, "dur": 0.398,
    "args": {
      "External id": 460,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 455, "finished": true, "Ev Idx": 459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564823684.303, "dur": 39.350,
    "args": {
      "External id": 461,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 455, "finished": true, "Ev Idx": 460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564823684.956, "dur": 2.364,
    "args": {
      "External id": 462,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 455, "finished": true, "Ev Idx": 461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823686.397, "dur": 0.391,
    "args": {
      "External id": 463,"Record function id": 0, "finished": true, "Ev Idx": 462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823687.807, "dur": 2.428,
    "args": {
      "External id": 464,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 456, "finished": true, "Ev Idx": 463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823688.187, "dur": 1.933,
    "args": {
      "External id": 465,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 456, "finished": true, "Ev Idx": 464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564823690.639, "dur": 1.211,
    "args": {
      "External id": 466,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 457, "finished": true, "Ev Idx": 465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823691.292, "dur": 0.254,
    "args": {
      "External id": 467,"Record function id": 0, "finished": true, "Ev Idx": 466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823692.165, "dur": 1.228,
    "args": {
      "External id": 468,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 458, "finished": true, "Ev Idx": 467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823692.463, "dur": 0.858,
    "args": {
      "External id": 469,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 458, "finished": true, "Ev Idx": 468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564823693.824, "dur": 27.790,
    "args": {
      "External id": 470,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 459, "finished": true, "Ev Idx": 469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564823704.233, "dur": 1.197,
    "args": {
      "External id": 471,"Record function id": 0, "finished": true, "Ev Idx": 470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564823722.169, "dur": 1.179,
    "args": {
      "External id": 472,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 460, "finished": true, "Ev Idx": 471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564823725.718, "dur": 1.906,
    "args": {
      "External id": 473,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 461, "finished": true, "Ev Idx": 472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823726.709, "dur": 0.375,
    "args": {
      "External id": 474,"Record function id": 0, "finished": true, "Ev Idx": 473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564823730.295, "dur": 13.275,
    "args": {
      "External id": 475,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 462, "finished": true, "Ev Idx": 474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564823730.800, "dur": 12.632,
    "args": {
      "External id": 476,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 462, "finished": true, "Ev Idx": 475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564823731.791, "dur": 1.929,
    "args": {
      "External id": 477,"Record function id": 0, "finished": true, "Ev Idx": 476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564823732.235, "dur": 1.326,
    "args": {
      "External id": 478,"Record function id": 0, "finished": true, "Ev Idx": 477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564823734.054, "dur": 8.931,
    "args": {
      "External id": 479,"Record function id": 0, "finished": true, "Ev Idx": 478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823749.136, "dur": 1.637,
    "args": {
      "External id": 480,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 463, "finished": true, "Ev Idx": 479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823749.455, "dur": 1.231,
    "args": {
      "External id": 481,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 463, "finished": true, "Ev Idx": 480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564823764.154, "dur": 36.815,
    "args": {
      "External id": 482,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 464, "finished": true, "Ev Idx": 481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564823764.963, "dur": 2.961,
    "args": {
      "External id": 483,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 464, "finished": true, "Ev Idx": 482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564823766.439, "dur": 0.949,
    "args": {
      "External id": 484,"Record function id": 0, "finished": true, "Ev Idx": 483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564823766.924, "dur": 0.349,
    "args": {
      "External id": 485,"Record function id": 0, "finished": true, "Ev Idx": 484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564823768.276, "dur": 32.540,
    "args": {
      "External id": 486,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 465, "finished": true, "Ev Idx": 485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564823768.674, "dur": 1.371,
    "args": {
      "External id": 487,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 465, "finished": true, "Ev Idx": 486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564823769.059, "dur": 0.901,
    "args": {
      "External id": 488,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 465, "finished": true, "Ev Idx": 487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564823770.596, "dur": 28.185,
    "args": {
      "External id": 489,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 466, "finished": true, "Ev Idx": 488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564823781.739, "dur": 1.149,
    "args": {
      "External id": 490,"Record function id": 0, "finished": true, "Ev Idx": 489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564823799.364, "dur": 1.077,
    "args": {
      "External id": 491,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 467, "finished": true, "Ev Idx": 490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564823805.726, "dur": 91.500,
    "args": {
      "External id": 492,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 468, "finished": true, "Ev Idx": 491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564823910.316, "dur": 0.243,
    "args": {
      "External id": 493,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 469, "finished": true, "Ev Idx": 492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564823912.965, "dur": 14.336,
    "args": {
      "External id": 494,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 469, "finished": true, "Ev Idx": 493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564823915.565, "dur": 0.489,
    "args": {
      "External id": 495,"Record function id": 0, "finished": true, "Ev Idx": 494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564823916.701, "dur": 0.221,
    "args": {
      "External id": 496,"Record function id": 0, "finished": true, "Ev Idx": 495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564823930.490, "dur": 15.525,
    "args": {
      "External id": 497,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 470, "finished": true, "Ev Idx": 496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564823950.247, "dur": 29.388,
    "args": {
      "External id": 498,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 471, "finished": true, "Ev Idx": 497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564823982.533, "dur": 11.703,
    "args": {
      "External id": 499,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 472, "finished": true, "Ev Idx": 498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564823996.188, "dur": 9.713,
    "args": {
      "External id": 500,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 473, "finished": true, "Ev Idx": 499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564824008.642, "dur": 0.112,
    "args": {
      "External id": 501,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 474, "finished": true, "Ev Idx": 500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564824009.817, "dur": 7.375,
    "args": {
      "External id": 502,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 474, "finished": true, "Ev Idx": 501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564824033.098, "dur": 65.426,
    "args": {
      "External id": 503,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 475, "finished": true, "Ev Idx": 502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564824033.860, "dur": 4.947,
    "args": {
      "External id": 504,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 475, "finished": true, "Ev Idx": 503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564824036.158, "dur": 1.740,
    "args": {
      "External id": 505,"Record function id": 0, "finished": true, "Ev Idx": 504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824037.099, "dur": 0.593,
    "args": {
      "External id": 506,"Record function id": 0, "finished": true, "Ev Idx": 505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564824039.157, "dur": 59.204,
    "args": {
      "External id": 507,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 476, "finished": true, "Ev Idx": 506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564824039.987, "dur": 2.035,
    "args": {
      "External id": 508,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 476, "finished": true, "Ev Idx": 507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564824040.325, "dur": 1.555,
    "args": {
      "External id": 509,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 476, "finished": true, "Ev Idx": 508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564824042.410, "dur": 51.251,
    "args": {
      "External id": 510,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 477, "finished": true, "Ev Idx": 509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564824055.328, "dur": 1.432,
    "args": {
      "External id": 511,"Record function id": 0, "finished": true, "Ev Idx": 510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564824095.473, "dur": 2.306,
    "args": {
      "External id": 512,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 478, "finished": true, "Ev Idx": 511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564824143.768, "dur": 14.119,
    "args": {
      "External id": 513,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 479, "finished": true, "Ev Idx": 512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564824166.229, "dur": 101.250,
    "args": {
      "External id": 514,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 480, "finished": true, "Ev Idx": 513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564824166.729, "dur": 3.051,
    "args": {
      "External id": 515,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 480, "finished": true, "Ev Idx": 514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564824167.912, "dur": 1.295,
    "args": {
      "External id": 516,"Record function id": 0, "finished": true, "Ev Idx": 515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824168.455, "dur": 0.565,
    "args": {
      "External id": 517,"Record function id": 0, "finished": true, "Ev Idx": 516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564824170.145, "dur": 97.162,
    "args": {
      "External id": 518,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 481, "finished": true, "Ev Idx": 517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564824170.663, "dur": 1.890,
    "args": {
      "External id": 519,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 481, "finished": true, "Ev Idx": 518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564824171.001, "dur": 1.439,
    "args": {
      "External id": 520,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 481, "finished": true, "Ev Idx": 519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564824172.947, "dur": 92.091,
    "args": {
      "External id": 521,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 482, "finished": true, "Ev Idx": 520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564824237.612, "dur": 1.308,
    "args": {
      "External id": 522,"Record function id": 0, "finished": true, "Ev Idx": 521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564824265.724, "dur": 1.233,
    "args": {
      "External id": 523,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 483, "finished": true, "Ev Idx": 522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564824270.289, "dur": 9.675,
    "args": {
      "External id": 524,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 484, "finished": true, "Ev Idx": 523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564824287.362, "dur": 41.577,
    "args": {
      "External id": 525,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 485, "finished": true, "Ev Idx": 524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564824287.925, "dur": 3.015,
    "args": {
      "External id": 526,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 485, "finished": true, "Ev Idx": 525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564824289.052, "dur": 1.195,
    "args": {
      "External id": 527,"Record function id": 0, "finished": true, "Ev Idx": 526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824289.644, "dur": 0.479,
    "args": {
      "External id": 528,"Record function id": 0, "finished": true, "Ev Idx": 527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564824291.352, "dur": 37.456,
    "args": {
      "External id": 529,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 486, "finished": true, "Ev Idx": 528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564824291.729, "dur": 1.569,
    "args": {
      "External id": 530,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 486, "finished": true, "Ev Idx": 529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564824292.082, "dur": 1.120,
    "args": {
      "External id": 531,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 486, "finished": true, "Ev Idx": 530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564824293.627, "dur": 33.356,
    "args": {
      "External id": 532,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 487, "finished": true, "Ev Idx": 531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564824305.520, "dur": 1.088,
    "args": {
      "External id": 533,"Record function id": 0, "finished": true, "Ev Idx": 532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564824327.587, "dur": 0.933,
    "args": {
      "External id": 534,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 488, "finished": true, "Ev Idx": 533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564824332.084, "dur": 8.492,
    "args": {
      "External id": 535,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 489, "finished": true, "Ev Idx": 534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564824362.736, "dur": 0.307,
    "args": {
      "External id": 536,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 490, "finished": true, "Ev Idx": 535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564824365.092, "dur": 10.462,
    "args": {
      "External id": 537,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 490, "finished": true, "Ev Idx": 536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564824367.108, "dur": 0.461,
    "args": {
      "External id": 538,"Record function id": 0, "finished": true, "Ev Idx": 537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564824367.922, "dur": 0.065,
    "args": {
      "External id": 539,"Record function id": 0, "finished": true, "Ev Idx": 538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564824378.315, "dur": 12.348,
    "args": {
      "External id": 540,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 491, "finished": true, "Ev Idx": 539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564824394.400, "dur": 8.617,
    "args": {
      "External id": 541,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 492, "finished": true, "Ev Idx": 540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564824404.713, "dur": 8.880,
    "args": {
      "External id": 542,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 493, "finished": true, "Ev Idx": 541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564824415.160, "dur": 8.320,
    "args": {
      "External id": 543,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 494, "finished": true, "Ev Idx": 542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564824426.246, "dur": 0.103,
    "args": {
      "External id": 544,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 495, "finished": true, "Ev Idx": 543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564824427.268, "dur": 51.424,
    "args": {
      "External id": 545,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 495, "finished": true, "Ev Idx": 544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564824506.101, "dur": 40.540,
    "args": {
      "External id": 546,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 496, "finished": true, "Ev Idx": 545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564824506.771, "dur": 3.348,
    "args": {
      "External id": 547,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 496, "finished": true, "Ev Idx": 546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564824507.929, "dur": 1.461,
    "args": {
      "External id": 548,"Record function id": 0, "finished": true, "Ev Idx": 547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824508.575, "dur": 0.602,
    "args": {
      "External id": 549,"Record function id": 0, "finished": true, "Ev Idx": 548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564824510.526, "dur": 36.036,
    "args": {
      "External id": 550,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 497, "finished": true, "Ev Idx": 549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564824511.145, "dur": 1.844,
    "args": {
      "External id": 551,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 497, "finished": true, "Ev Idx": 550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564824511.474, "dur": 1.365,
    "args": {
      "External id": 552,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 497, "finished": true, "Ev Idx": 551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564824513.329, "dur": 30.981,
    "args": {
      "External id": 553,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 498, "finished": true, "Ev Idx": 552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564824525.058, "dur": 1.372,
    "args": {
      "External id": 554,"Record function id": 0, "finished": true, "Ev Idx": 553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564824545.095, "dur": 1.136,
    "args": {
      "External id": 555,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 499, "finished": true, "Ev Idx": 554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564824550.371, "dur": 1.684,
    "args": {
      "External id": 556,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 500, "finished": true, "Ev Idx": 555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564824554.155, "dur": 2.686,
    "args": {
      "External id": 557,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 501, "finished": true, "Ev Idx": 556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824555.430, "dur": 0.564,
    "args": {
      "External id": 558,"Record function id": 0, "finished": true, "Ev Idx": 557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564824564.979, "dur": 41.359,
    "args": {
      "External id": 559,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 502, "finished": true, "Ev Idx": 558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564824565.311, "dur": 2.307,
    "args": {
      "External id": 560,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 502, "finished": true, "Ev Idx": 559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564824566.142, "dur": 1.039,
    "args": {
      "External id": 561,"Record function id": 0, "finished": true, "Ev Idx": 560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824566.754, "dur": 0.296,
    "args": {
      "External id": 562,"Record function id": 0, "finished": true, "Ev Idx": 561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564824567.954, "dur": 38.286,
    "args": {
      "External id": 563,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 503, "finished": true, "Ev Idx": 562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564824568.390, "dur": 1.281,
    "args": {
      "External id": 564,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 503, "finished": true, "Ev Idx": 563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564824568.672, "dur": 0.913,
    "args": {
      "External id": 565,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 503, "finished": true, "Ev Idx": 564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564824569.997, "dur": 34.366,
    "args": {
      "External id": 566,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 504, "finished": true, "Ev Idx": 565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564824580.184, "dur": 1.120,
    "args": {
      "External id": 567,"Record function id": 0, "finished": true, "Ev Idx": 566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564824604.931, "dur": 1.026,
    "args": {
      "External id": 568,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 505, "finished": true, "Ev Idx": 567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564824608.893, "dur": 1.240,
    "args": {
      "External id": 569,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 506, "finished": true, "Ev Idx": 568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564824611.702, "dur": 1.765,
    "args": {
      "External id": 570,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 507, "finished": true, "Ev Idx": 569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824612.590, "dur": 0.364,
    "args": {
      "External id": 571,"Record function id": 0, "finished": true, "Ev Idx": 570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564824621.342, "dur": 35.923,
    "args": {
      "External id": 572,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 508, "finished": true, "Ev Idx": 571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564824621.693, "dur": 1.964,
    "args": {
      "External id": 573,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 508, "finished": true, "Ev Idx": 572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564824622.438, "dur": 0.813,
    "args": {
      "External id": 574,"Record function id": 0, "finished": true, "Ev Idx": 573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824622.875, "dur": 0.264,
    "args": {
      "External id": 575,"Record function id": 0, "finished": true, "Ev Idx": 574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564824623.967, "dur": 33.200,
    "args": {
      "External id": 576,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 509, "finished": true, "Ev Idx": 575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564824624.326, "dur": 1.190,
    "args": {
      "External id": 577,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 509, "finished": true, "Ev Idx": 576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564824624.585, "dur": 0.828,
    "args": {
      "External id": 578,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 509, "finished": true, "Ev Idx": 577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564824625.816, "dur": 29.718,
    "args": {
      "External id": 579,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 510, "finished": true, "Ev Idx": 578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564824635.574, "dur": 1.048,
    "args": {
      "External id": 580,"Record function id": 0, "finished": true, "Ev Idx": 579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564824656.021, "dur": 0.890,
    "args": {
      "External id": 581,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 511, "finished": true, "Ev Idx": 580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564824659.854, "dur": 1.108,
    "args": {
      "External id": 582,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 512, "finished": true, "Ev Idx": 581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564824662.352, "dur": 2.063,
    "args": {
      "External id": 583,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 513, "finished": true, "Ev Idx": 582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824663.073, "dur": 0.652,
    "args": {
      "External id": 584,"Record function id": 0, "finished": true, "Ev Idx": 583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564824667.824, "dur": 2.471,
    "args": {
      "External id": 585,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 514, "finished": true, "Ev Idx": 584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824669.444, "dur": 0.308,
    "args": {
      "External id": 586,"Record function id": 0, "finished": true, "Ev Idx": 585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564824671.460, "dur": 1.105,
    "args": {
      "External id": 587,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 514, "finished": true, "Ev Idx": 586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824671.927, "dur": 0.281,
    "args": {
      "External id": 588,"Record function id": 0, "finished": true, "Ev Idx": 587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564824673.615, "dur": 10.304,
    "args": {
      "External id": 589,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 514, "finished": true, "Ev Idx": 588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824690.610, "dur": 6.092,
    "args": {
      "External id": 590,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 515, "finished": true, "Ev Idx": 589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824694.150, "dur": 0.923,
    "args": {
      "External id": 591,"Record function id": 0, "finished": true, "Ev Idx": 590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824699.088, "dur": 3.166,
    "args": {
      "External id": 592,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 516, "finished": true, "Ev Idx": 591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824700.877, "dur": 0.573,
    "args": {
      "External id": 593,"Record function id": 0, "finished": true, "Ev Idx": 592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564824704.391, "dur": 11.118,
    "args": {
      "External id": 594,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 517, "finished": true, "Ev Idx": 593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564824720.476, "dur": 19.419,
    "args": {
      "External id": 595,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 518, "finished": true, "Ev Idx": 594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564824742.485, "dur": 9.082,
    "args": {
      "External id": 596,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 519, "finished": true, "Ev Idx": 595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564824753.236, "dur": 8.659,
    "args": {
      "External id": 597,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 520, "finished": true, "Ev Idx": 596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564824763.986, "dur": 7.582,
    "args": {
      "External id": 598,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 521, "finished": true, "Ev Idx": 597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824774.173, "dur": 2.980,
    "args": {
      "External id": 599,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 522, "finished": true, "Ev Idx": 598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824776.029, "dur": 0.388,
    "args": {
      "External id": 600,"Record function id": 0, "finished": true, "Ev Idx": 599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824778.742, "dur": 2.695,
    "args": {
      "External id": 601,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 523, "finished": true, "Ev Idx": 600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824780.405, "dur": 0.457,
    "args": {
      "External id": 602,"Record function id": 0, "finished": true, "Ev Idx": 601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564824782.485, "dur": 7.501,
    "args": {
      "External id": 603,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 524, "finished": true, "Ev Idx": 602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564824791.682, "dur": 11.809,
    "args": {
      "External id": 604,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 525, "finished": true, "Ev Idx": 603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564824805.718, "dur": 6.343,
    "args": {
      "External id": 605,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 526, "finished": true, "Ev Idx": 604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564824813.634, "dur": 6.456,
    "args": {
      "External id": 606,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 527, "finished": true, "Ev Idx": 605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824845.684, "dur": 3.658,
    "args": {
      "External id": 607,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 528, "finished": true, "Ev Idx": 606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824847.939, "dur": 0.628,
    "args": {
      "External id": 608,"Record function id": 0, "finished": true, "Ev Idx": 607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824849.996, "dur": 2.462,
    "args": {
      "External id": 609,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 529, "finished": true, "Ev Idx": 608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824851.453, "dur": 0.452,
    "args": {
      "External id": 610,"Record function id": 0, "finished": true, "Ev Idx": 609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564824853.259, "dur": 2.627,
    "args": {
      "External id": 611,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 530, "finished": true, "Ev Idx": 610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824854.397, "dur": 0.793,
    "args": {
      "External id": 612,"Record function id": 0, "finished": true, "Ev Idx": 611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824856.534, "dur": 1.732,
    "args": {
      "External id": 613,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 531, "finished": true, "Ev Idx": 612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824857.319, "dur": 0.402,
    "args": {
      "External id": 614,"Record function id": 0, "finished": true, "Ev Idx": 613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824858.897, "dur": 1.499,
    "args": {
      "External id": 615,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 532, "finished": true, "Ev Idx": 614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824859.609, "dur": 0.334,
    "args": {
      "External id": 616,"Record function id": 0, "finished": true, "Ev Idx": 615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564824863.548, "dur": 3.307,
    "args": {
      "External id": 617,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 533, "finished": true, "Ev Idx": 616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824865.648, "dur": 0.405,
    "args": {
      "External id": 618,"Record function id": 0, "finished": true, "Ev Idx": 617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564824869.558, "dur": 22.229,
    "args": {
      "External id": 619,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 534, "finished": true, "Ev Idx": 618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564824872.116, "dur": 17.729,
    "args": {
      "External id": 620,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 534, "finished": true, "Ev Idx": 619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564824875.222, "dur": 2.840,
    "args": {
      "External id": 621,"Record function id": 0, "finished": true, "Ev Idx": 620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564824876.122, "dur": 1.740,
    "args": {
      "External id": 622,"Record function id": 0, "finished": true, "Ev Idx": 621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564824878.946, "dur": 10.337,
    "args": {
      "External id": 623,"Record function id": 0, "finished": true, "Ev Idx": 622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564824890.358, "dur": 1.188,
    "args": {
      "External id": 624,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 535, "finished": true, "Ev Idx": 623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824895.396, "dur": 3.039,
    "args": {
      "External id": 625,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 536, "finished": true, "Ev Idx": 624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824897.487, "dur": 0.397,
    "args": {
      "External id": 626,"Record function id": 0, "finished": true, "Ev Idx": 625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824899.180, "dur": 2.866,
    "args": {
      "External id": 627,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 537, "finished": true, "Ev Idx": 626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824901.259, "dur": 0.287,
    "args": {
      "External id": 628,"Record function id": 0, "finished": true, "Ev Idx": 627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564824902.590, "dur": 2.026,
    "args": {
      "External id": 629,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 538, "finished": true, "Ev Idx": 628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824903.833, "dur": 0.347,
    "args": {
      "External id": 630,"Record function id": 0, "finished": true, "Ev Idx": 629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824905.188, "dur": 2.634,
    "args": {
      "External id": 631,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 539, "finished": true, "Ev Idx": 630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824907.000, "dur": 0.324,
    "args": {
      "External id": 632,"Record function id": 0, "finished": true, "Ev Idx": 631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824908.429, "dur": 2.410,
    "args": {
      "External id": 633,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 540, "finished": true, "Ev Idx": 632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824909.958, "dur": 0.427,
    "args": {
      "External id": 634,"Record function id": 0, "finished": true, "Ev Idx": 633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564824912.474, "dur": 1.883,
    "args": {
      "External id": 635,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 541, "finished": true, "Ev Idx": 634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824913.452, "dur": 0.477,
    "args": {
      "External id": 636,"Record function id": 0, "finished": true, "Ev Idx": 635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564824915.765, "dur": 11.118,
    "args": {
      "External id": 637,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 542, "finished": true, "Ev Idx": 636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564824916.695, "dur": 8.691,
    "args": {
      "External id": 638,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 542, "finished": true, "Ev Idx": 637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564824917.327, "dur": 1.719,
    "args": {
      "External id": 639,"Record function id": 0, "finished": true, "Ev Idx": 638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564824917.679, "dur": 1.249,
    "args": {
      "External id": 640,"Record function id": 0, "finished": true, "Ev Idx": 639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564824919.390, "dur": 5.610,
    "args": {
      "External id": 641,"Record function id": 0, "finished": true, "Ev Idx": 640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564824925.775, "dur": 0.907,
    "args": {
      "External id": 642,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 543, "finished": true, "Ev Idx": 641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564824929.012, "dur": 1.953,
    "args": {
      "External id": 643,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 544, "finished": true, "Ev Idx": 642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824930.127, "dur": 0.363,
    "args": {
      "External id": 644,"Record function id": 0, "finished": true, "Ev Idx": 643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564824932.564, "dur": 45.176,
    "args": {
      "External id": 645,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 545, "finished": true, "Ev Idx": 644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564824933.253, "dur": 2.103,
    "args": {
      "External id": 646,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 545, "finished": true, "Ev Idx": 645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824934.625, "dur": 0.312,
    "args": {
      "External id": 647,"Record function id": 0, "finished": true, "Ev Idx": 646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564824935.718, "dur": 4.091,
    "args": {
      "External id": 648,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 546, "finished": true, "Ev Idx": 647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564824937.741, "dur": 1.935,
    "args": {
      "External id": 649,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 546, "finished": true, "Ev Idx": 648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564824940.296, "dur": 1.268,
    "args": {
      "External id": 650,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 547, "finished": true, "Ev Idx": 649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824940.927, "dur": 0.252,
    "args": {
      "External id": 651,"Record function id": 0, "finished": true, "Ev Idx": 650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564824941.962, "dur": 1.790,
    "args": {
      "External id": 652,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 548, "finished": true, "Ev Idx": 651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564824942.792, "dur": 0.852,
    "args": {
      "External id": 653,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 548, "finished": true, "Ev Idx": 652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564824944.637, "dur": 31.098,
    "args": {
      "External id": 654,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 549, "finished": true, "Ev Idx": 653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564824957.785, "dur": 0.938,
    "args": {
      "External id": 655,"Record function id": 0, "finished": true, "Ev Idx": 654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564824976.340, "dur": 1.055,
    "args": {
      "External id": 656,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 550, "finished": true, "Ev Idx": 655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564824981.216, "dur": 9.647,
    "args": {
      "External id": 657,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 551, "finished": true, "Ev Idx": 656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824994.005, "dur": 1.968,
    "args": {
      "External id": 658,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 552, "finished": true, "Ev Idx": 657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824994.796, "dur": 0.537,
    "args": {
      "External id": 659,"Record function id": 0, "finished": true, "Ev Idx": 658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824996.815, "dur": 1.312,
    "args": {
      "External id": 660,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 552, "finished": true, "Ev Idx": 659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824997.448, "dur": 0.302,
    "args": {
      "External id": 661,"Record function id": 0, "finished": true, "Ev Idx": 660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564824998.755, "dur": 1.043,
    "args": {
      "External id": 662,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 552, "finished": true, "Ev Idx": 661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564824999.209, "dur": 0.256,
    "args": {
      "External id": 663,"Record function id": 0, "finished": true, "Ev Idx": 662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564825001.447, "dur": 7.314,
    "args": {
      "External id": 664,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 552, "finished": true, "Ev Idx": 663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564825014.793, "dur": 14.167,
    "args": {
      "External id": 665,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 553, "finished": true, "Ev Idx": 664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564825015.783, "dur": 0.589,
    "args": {
      "External id": 666,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 553, "finished": true, "Ev Idx": 665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564825017.119, "dur": 11.534,
    "args": {
      "External id": 667,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 553, "finished": true, "Ev Idx": 666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564825031.950, "dur": 0.226,
    "args": {
      "External id": 668,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 554, "finished": true, "Ev Idx": 667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564825044.499, "dur": 0.647,
    "args": {
      "External id": 669,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 554, "finished": true, "Ev Idx": 668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564825046.833, "dur": 42.762,
    "args": {
      "External id": 670,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 554, "finished": true, "Ev Idx": 669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564825047.312, "dur": 2.208,
    "args": {
      "External id": 671,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 554, "finished": true, "Ev Idx": 670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825048.388, "dur": 0.656,
    "args": {
      "External id": 672,"Record function id": 0, "finished": true, "Ev Idx": 671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564825049.863, "dur": 2.806,
    "args": {
      "External id": 673,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 555, "finished": true, "Ev Idx": 672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825050.405, "dur": 2.148,
    "args": {
      "External id": 674,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 555, "finished": true, "Ev Idx": 673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564825053.176, "dur": 2.279,
    "args": {
      "External id": 675,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 556, "finished": true, "Ev Idx": 674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825054.802, "dur": 0.277,
    "args": {
      "External id": 676,"Record function id": 0, "finished": true, "Ev Idx": 675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564825055.715, "dur": 2.360,
    "args": {
      "External id": 677,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 557, "finished": true, "Ev Idx": 676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825056.137, "dur": 1.861,
    "args": {
      "External id": 678,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 557, "finished": true, "Ev Idx": 677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564825058.432, "dur": 29.266,
    "args": {
      "External id": 679,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 558, "finished": true, "Ev Idx": 678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564825071.827, "dur": 1.090,
    "args": {
      "External id": 680,"Record function id": 0, "finished": true, "Ev Idx": 679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564825088.249, "dur": 1.052,
    "args": {
      "External id": 681,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 559, "finished": true, "Ev Idx": 680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825091.549, "dur": 1.788,
    "args": {
      "External id": 682,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 560, "finished": true, "Ev Idx": 681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825092.377, "dur": 0.413,
    "args": {
      "External id": 683,"Record function id": 0, "finished": true, "Ev Idx": 682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564825095.909, "dur": 11.593,
    "args": {
      "External id": 684,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 561, "finished": true, "Ev Idx": 683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564825096.460, "dur": 10.904,
    "args": {
      "External id": 685,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 561, "finished": true, "Ev Idx": 684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564825097.539, "dur": 1.846,
    "args": {
      "External id": 686,"Record function id": 0, "finished": true, "Ev Idx": 685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564825097.929, "dur": 1.324,
    "args": {
      "External id": 687,"Record function id": 0, "finished": true, "Ev Idx": 686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564825099.871, "dur": 7.065,
    "args": {
      "External id": 688,"Record function id": 0, "finished": true, "Ev Idx": 687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564825113.269, "dur": 1.867,
    "args": {
      "External id": 689,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 562, "finished": true, "Ev Idx": 688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825113.733, "dur": 1.301,
    "args": {
      "External id": 690,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 562, "finished": true, "Ev Idx": 689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564825127.680, "dur": 35.094,
    "args": {
      "External id": 691,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 563, "finished": true, "Ev Idx": 690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564825128.341, "dur": 3.096,
    "args": {
      "External id": 692,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 563, "finished": true, "Ev Idx": 691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825129.701, "dur": 1.127,
    "args": {
      "External id": 693,"Record function id": 0, "finished": true, "Ev Idx": 692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825130.359, "dur": 0.346,
    "args": {
      "External id": 694,"Record function id": 0, "finished": true, "Ev Idx": 693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564825131.804, "dur": 30.831,
    "args": {
      "External id": 695,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 564, "finished": true, "Ev Idx": 694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564825132.262, "dur": 2.418,
    "args": {
      "External id": 696,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 564, "finished": true, "Ev Idx": 695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825132.535, "dur": 2.036,
    "args": {
      "External id": 697,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 564, "finished": true, "Ev Idx": 696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564825135.051, "dur": 25.738,
    "args": {
      "External id": 698,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 565, "finished": true, "Ev Idx": 697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564825145.780, "dur": 1.148,
    "args": {
      "External id": 699,"Record function id": 0, "finished": true, "Ev Idx": 698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564825161.338, "dur": 0.987,
    "args": {
      "External id": 700,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 566, "finished": true, "Ev Idx": 699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564825167.153, "dur": 8.933,
    "args": {
      "External id": 701,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 567, "finished": true, "Ev Idx": 700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564825184.821, "dur": 0.209,
    "args": {
      "External id": 702,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 568, "finished": true, "Ev Idx": 701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564825186.931, "dur": 11.648,
    "args": {
      "External id": 703,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 568, "finished": true, "Ev Idx": 702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564825189.159, "dur": 0.406,
    "args": {
      "External id": 704,"Record function id": 0, "finished": true, "Ev Idx": 703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564825190.166, "dur": 0.077,
    "args": {
      "External id": 705,"Record function id": 0, "finished": true, "Ev Idx": 704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564825201.491, "dur": 12.545,
    "args": {
      "External id": 706,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 569, "finished": true, "Ev Idx": 705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564825217.979, "dur": 8.356,
    "args": {
      "External id": 707,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 570, "finished": true, "Ev Idx": 706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564825228.275, "dur": 10.110,
    "args": {
      "External id": 708,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 571, "finished": true, "Ev Idx": 707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564825240.156, "dur": 9.222,
    "args": {
      "External id": 709,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 572, "finished": true, "Ev Idx": 708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564825252.006, "dur": 0.299,
    "args": {
      "External id": 710,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 573, "finished": true, "Ev Idx": 709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564825253.368, "dur": 8.003,
    "args": {
      "External id": 711,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 573, "finished": true, "Ev Idx": 710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564825276.393, "dur": 117.501,
    "args": {
      "External id": 712,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 574, "finished": true, "Ev Idx": 711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564825277.153, "dur": 4.001,
    "args": {
      "External id": 713,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 574, "finished": true, "Ev Idx": 712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825279.300, "dur": 1.128,
    "args": {
      "External id": 714,"Record function id": 0, "finished": true, "Ev Idx": 713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825279.822, "dur": 0.421,
    "args": {
      "External id": 715,"Record function id": 0, "finished": true, "Ev Idx": 714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564825281.527, "dur": 112.205,
    "args": {
      "External id": 716,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 575, "finished": true, "Ev Idx": 715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564825282.071, "dur": 2.181,
    "args": {
      "External id": 717,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 575, "finished": true, "Ev Idx": 716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825282.550, "dur": 1.572,
    "args": {
      "External id": 718,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 575, "finished": true, "Ev Idx": 717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564825284.567, "dur": 106.571,
    "args": {
      "External id": 719,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 576, "finished": true, "Ev Idx": 718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564825358.874, "dur": 1.307,
    "args": {
      "External id": 720,"Record function id": 0, "finished": true, "Ev Idx": 719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564825391.983, "dur": 1.371,
    "args": {
      "External id": 721,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 577, "finished": true, "Ev Idx": 720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564825404.303, "dur": 11.553,
    "args": {
      "External id": 722,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 578, "finished": true, "Ev Idx": 721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564825424.814, "dur": 83.283,
    "args": {
      "External id": 723,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 579, "finished": true, "Ev Idx": 722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564825425.347, "dur": 4.649,
    "args": {
      "External id": 724,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 579, "finished": true, "Ev Idx": 723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825427.576, "dur": 1.408,
    "args": {
      "External id": 725,"Record function id": 0, "finished": true, "Ev Idx": 724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825428.221, "dur": 0.643,
    "args": {
      "External id": 726,"Record function id": 0, "finished": true, "Ev Idx": 725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564825430.402, "dur": 77.573,
    "args": {
      "External id": 727,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 580, "finished": true, "Ev Idx": 726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564825431.015, "dur": 1.907,
    "args": {
      "External id": 728,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 580, "finished": true, "Ev Idx": 727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825431.559, "dur": 1.248,
    "args": {
      "External id": 729,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 580, "finished": true, "Ev Idx": 728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564825433.300, "dur": 72.450,
    "args": {
      "External id": 730,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 581, "finished": true, "Ev Idx": 729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564825479.745, "dur": 1.266,
    "args": {
      "External id": 731,"Record function id": 0, "finished": true, "Ev Idx": 730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564825506.469, "dur": 1.186,
    "args": {
      "External id": 732,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 582, "finished": true, "Ev Idx": 731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564825510.594, "dur": 9.371,
    "args": {
      "External id": 733,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 583, "finished": true, "Ev Idx": 732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564825527.450, "dur": 41.025,
    "args": {
      "External id": 734,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 584, "finished": true, "Ev Idx": 733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564825527.914, "dur": 2.780,
    "args": {
      "External id": 735,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 584, "finished": true, "Ev Idx": 734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825528.932, "dur": 1.157,
    "args": {
      "External id": 736,"Record function id": 0, "finished": true, "Ev Idx": 735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825529.475, "dur": 0.493,
    "args": {
      "External id": 737,"Record function id": 0, "finished": true, "Ev Idx": 736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564825531.043, "dur": 37.321,
    "args": {
      "External id": 738,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 585, "finished": true, "Ev Idx": 737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564825531.660, "dur": 3.520,
    "args": {
      "External id": 739,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 585, "finished": true, "Ev Idx": 738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825532.037, "dur": 3.028,
    "args": {
      "External id": 740,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 585, "finished": true, "Ev Idx": 739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564825535.524, "dur": 30.856,
    "args": {
      "External id": 741,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 586, "finished": true, "Ev Idx": 740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564825544.357, "dur": 1.145,
    "args": {
      "External id": 742,"Record function id": 0, "finished": true, "Ev Idx": 741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564825566.978, "dur": 1.073,
    "args": {
      "External id": 743,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 587, "finished": true, "Ev Idx": 742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564825571.355, "dur": 9.721,
    "args": {
      "External id": 744,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 588, "finished": true, "Ev Idx": 743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564825602.454, "dur": 0.359,
    "args": {
      "External id": 745,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 589, "finished": true, "Ev Idx": 744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564825605.002, "dur": 10.995,
    "args": {
      "External id": 746,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 589, "finished": true, "Ev Idx": 745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564825606.479, "dur": 0.486,
    "args": {
      "External id": 747,"Record function id": 0, "finished": true, "Ev Idx": 746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564825607.347, "dur": 0.240,
    "args": {
      "External id": 748,"Record function id": 0, "finished": true, "Ev Idx": 747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564825618.942, "dur": 12.783,
    "args": {
      "External id": 749,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 590, "finished": true, "Ev Idx": 748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564825635.698, "dur": 8.386,
    "args": {
      "External id": 750,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 591, "finished": true, "Ev Idx": 749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564825645.793, "dur": 10.592,
    "args": {
      "External id": 751,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 592, "finished": true, "Ev Idx": 750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564825658.118, "dur": 10.675,
    "args": {
      "External id": 752,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 593, "finished": true, "Ev Idx": 751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564825671.388, "dur": 0.260,
    "args": {
      "External id": 753,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 594, "finished": true, "Ev Idx": 752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564825672.463, "dur": 7.349,
    "args": {
      "External id": 754,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 594, "finished": true, "Ev Idx": 753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564825702.558, "dur": 35.567,
    "args": {
      "External id": 755,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 595, "finished": true, "Ev Idx": 754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564825703.382, "dur": 3.259,
    "args": {
      "External id": 756,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 595, "finished": true, "Ev Idx": 755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825704.546, "dur": 1.528,
    "args": {
      "External id": 757,"Record function id": 0, "finished": true, "Ev Idx": 756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825705.293, "dur": 0.643,
    "args": {
      "External id": 758,"Record function id": 0, "finished": true, "Ev Idx": 757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564825706.968, "dur": 31.041,
    "args": {
      "External id": 759,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 596, "finished": true, "Ev Idx": 758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564825707.549, "dur": 1.876,
    "args": {
      "External id": 760,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 596, "finished": true, "Ev Idx": 759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825708.250, "dur": 1.073,
    "args": {
      "External id": 761,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 596, "finished": true, "Ev Idx": 760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564825709.839, "dur": 26.161,
    "args": {
      "External id": 762,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 597, "finished": true, "Ev Idx": 761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564825718.636, "dur": 1.265,
    "args": {
      "External id": 763,"Record function id": 0, "finished": true, "Ev Idx": 762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564825736.667, "dur": 1.025,
    "args": {
      "External id": 764,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 598, "finished": true, "Ev Idx": 763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825741.425, "dur": 1.562,
    "args": {
      "External id": 765,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 599, "finished": true, "Ev Idx": 764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825745.323, "dur": 2.760,
    "args": {
      "External id": 766,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 600, "finished": true, "Ev Idx": 765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825746.347, "dur": 0.609,
    "args": {
      "External id": 767,"Record function id": 0, "finished": true, "Ev Idx": 766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564825754.924, "dur": 40.127,
    "args": {
      "External id": 768,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 601, "finished": true, "Ev Idx": 767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564825755.241, "dur": 2.321,
    "args": {
      "External id": 769,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 601, "finished": true, "Ev Idx": 768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825756.191, "dur": 0.903,
    "args": {
      "External id": 770,"Record function id": 0, "finished": true, "Ev Idx": 769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825756.696, "dur": 0.307,
    "args": {
      "External id": 771,"Record function id": 0, "finished": true, "Ev Idx": 770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564825757.883, "dur": 37.070,
    "args": {
      "External id": 772,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 602, "finished": true, "Ev Idx": 771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564825758.431, "dur": 1.298,
    "args": {
      "External id": 773,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 602, "finished": true, "Ev Idx": 772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825758.714, "dur": 0.921,
    "args": {
      "External id": 774,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 602, "finished": true, "Ev Idx": 773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564825760.012, "dur": 33.007,
    "args": {
      "External id": 775,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 603, "finished": true, "Ev Idx": 774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564825771.020, "dur": 1.100,
    "args": {
      "External id": 776,"Record function id": 0, "finished": true, "Ev Idx": 775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564825793.569, "dur": 1.096,
    "args": {
      "External id": 777,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 604, "finished": true, "Ev Idx": 776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825797.290, "dur": 1.279,
    "args": {
      "External id": 778,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 605, "finished": true, "Ev Idx": 777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825800.048, "dur": 1.917,
    "args": {
      "External id": 779,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 606, "finished": true, "Ev Idx": 778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825800.825, "dur": 0.525,
    "args": {
      "External id": 780,"Record function id": 0, "finished": true, "Ev Idx": 779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564825809.257, "dur": 33.403,
    "args": {
      "External id": 781,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 607, "finished": true, "Ev Idx": 780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564825809.532, "dur": 2.138,
    "args": {
      "External id": 782,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 607, "finished": true, "Ev Idx": 781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825810.294, "dur": 0.961,
    "args": {
      "External id": 783,"Record function id": 0, "finished": true, "Ev Idx": 782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825810.789, "dur": 0.372,
    "args": {
      "External id": 784,"Record function id": 0, "finished": true, "Ev Idx": 783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564825811.987, "dur": 30.574,
    "args": {
      "External id": 785,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 608, "finished": true, "Ev Idx": 784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564825812.605, "dur": 1.770,
    "args": {
      "External id": 786,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 608, "finished": true, "Ev Idx": 785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825813.051, "dur": 1.220,
    "args": {
      "External id": 787,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 608, "finished": true, "Ev Idx": 786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564825814.798, "dur": 26.015,
    "args": {
      "External id": 788,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 609, "finished": true, "Ev Idx": 787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564825822.333, "dur": 0.986,
    "args": {
      "External id": 789,"Record function id": 0, "finished": true, "Ev Idx": 788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564825841.340, "dur": 0.950,
    "args": {
      "External id": 790,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 610, "finished": true, "Ev Idx": 789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564825845.007, "dur": 1.149,
    "args": {
      "External id": 791,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 611, "finished": true, "Ev Idx": 790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564825847.586, "dur": 1.686,
    "args": {
      "External id": 792,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 612, "finished": true, "Ev Idx": 791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825848.372, "dur": 0.381,
    "args": {
      "External id": 793,"Record function id": 0, "finished": true, "Ev Idx": 792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564825852.131, "dur": 2.089,
    "args": {
      "External id": 794,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 613, "finished": true, "Ev Idx": 793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825853.333, "dur": 0.300,
    "args": {
      "External id": 795,"Record function id": 0, "finished": true, "Ev Idx": 794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564825855.422, "dur": 1.067,
    "args": {
      "External id": 796,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 613, "finished": true, "Ev Idx": 795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825855.864, "dur": 0.261,
    "args": {
      "External id": 797,"Record function id": 0, "finished": true, "Ev Idx": 796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564825857.364, "dur": 9.748,
    "args": {
      "External id": 798,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 613, "finished": true, "Ev Idx": 797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564825872.760, "dur": 5.720,
    "args": {
      "External id": 799,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 614, "finished": true, "Ev Idx": 798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825875.993, "dur": 0.895,
    "args": {
      "External id": 800,"Record function id": 0, "finished": true, "Ev Idx": 799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564825880.717, "dur": 3.219,
    "args": {
      "External id": 801,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 615, "finished": true, "Ev Idx": 800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825882.503, "dur": 0.680,
    "args": {
      "External id": 802,"Record function id": 0, "finished": true, "Ev Idx": 801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564825885.971, "dur": 11.197,
    "args": {
      "External id": 803,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 616, "finished": true, "Ev Idx": 802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564825901.742, "dur": 59.461,
    "args": {
      "External id": 804,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 617, "finished": true, "Ev Idx": 803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564825964.378, "dur": 10.117,
    "args": {
      "External id": 805,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 618, "finished": true, "Ev Idx": 804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564825976.517, "dur": 8.879,
    "args": {
      "External id": 806,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 619, "finished": true, "Ev Idx": 805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564825987.747, "dur": 7.699,
    "args": {
      "External id": 807,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 620, "finished": true, "Ev Idx": 806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564825998.474, "dur": 2.959,
    "args": {
      "External id": 808,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 621, "finished": true, "Ev Idx": 807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564825999.964, "dur": 0.598,
    "args": {
      "External id": 809,"Record function id": 0, "finished": true, "Ev Idx": 808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826003.149, "dur": 2.078,
    "args": {
      "External id": 810,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 622, "finished": true, "Ev Idx": 809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826004.139, "dur": 0.561,
    "args": {
      "External id": 811,"Record function id": 0, "finished": true, "Ev Idx": 810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564826006.521, "dur": 8.478,
    "args": {
      "External id": 812,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 623, "finished": true, "Ev Idx": 811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564826017.465, "dur": 10.588,
    "args": {
      "External id": 813,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 624, "finished": true, "Ev Idx": 812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564826030.402, "dur": 7.695,
    "args": {
      "External id": 814,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 625, "finished": true, "Ev Idx": 813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564826039.479, "dur": 8.247,
    "args": {
      "External id": 815,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 626, "finished": true, "Ev Idx": 814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826091.683, "dur": 4.649,
    "args": {
      "External id": 816,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 627, "finished": true, "Ev Idx": 815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826093.945, "dur": 0.893,
    "args": {
      "External id": 817,"Record function id": 0, "finished": true, "Ev Idx": 816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826096.971, "dur": 2.175,
    "args": {
      "External id": 818,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 628, "finished": true, "Ev Idx": 817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826098.328, "dur": 0.334,
    "args": {
      "External id": 819,"Record function id": 0, "finished": true, "Ev Idx": 818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564826100.032, "dur": 3.028,
    "args": {
      "External id": 820,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 629, "finished": true, "Ev Idx": 819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826101.503, "dur": 0.864,
    "args": {
      "External id": 821,"Record function id": 0, "finished": true, "Ev Idx": 820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826103.940, "dur": 1.697,
    "args": {
      "External id": 822,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 630, "finished": true, "Ev Idx": 821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826104.734, "dur": 0.404,
    "args": {
      "External id": 823,"Record function id": 0, "finished": true, "Ev Idx": 822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826106.293, "dur": 1.619,
    "args": {
      "External id": 824,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 631, "finished": true, "Ev Idx": 823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826106.980, "dur": 0.487,
    "args": {
      "External id": 825,"Record function id": 0, "finished": true, "Ev Idx": 824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564826111.580, "dur": 3.762,
    "args": {
      "External id": 826,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 632, "finished": true, "Ev Idx": 825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826114.108, "dur": 0.415,
    "args": {
      "External id": 827,"Record function id": 0, "finished": true, "Ev Idx": 826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826118.219, "dur": 25.627,
    "args": {
      "External id": 828,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 633, "finished": true, "Ev Idx": 827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564826121.221, "dur": 20.001,
    "args": {
      "External id": 829,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 633, "finished": true, "Ev Idx": 828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564826123.618, "dur": 3.919,
    "args": {
      "External id": 830,"Record function id": 0, "finished": true, "Ev Idx": 829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564826124.615, "dur": 2.632,
    "args": {
      "External id": 831,"Record function id": 0, "finished": true, "Ev Idx": 830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564826128.594, "dur": 11.933,
    "args": {
      "External id": 832,"Record function id": 0, "finished": true, "Ev Idx": 831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564826141.909, "dur": 1.611,
    "args": {
      "External id": 833,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 634, "finished": true, "Ev Idx": 832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826147.427, "dur": 2.721,
    "args": {
      "External id": 834,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 635, "finished": true, "Ev Idx": 833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826149.142, "dur": 0.411,
    "args": {
      "External id": 835,"Record function id": 0, "finished": true, "Ev Idx": 834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826150.793, "dur": 1.613,
    "args": {
      "External id": 836,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 636, "finished": true, "Ev Idx": 835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826151.689, "dur": 0.281,
    "args": {
      "External id": 837,"Record function id": 0, "finished": true, "Ev Idx": 836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564826152.961, "dur": 3.123,
    "args": {
      "External id": 838,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 637, "finished": true, "Ev Idx": 837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826155.052, "dur": 0.583,
    "args": {
      "External id": 839,"Record function id": 0, "finished": true, "Ev Idx": 838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826156.840, "dur": 2.354,
    "args": {
      "External id": 840,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 638, "finished": true, "Ev Idx": 839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826158.434, "dur": 0.294,
    "args": {
      "External id": 841,"Record function id": 0, "finished": true, "Ev Idx": 840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826159.937, "dur": 1.566,
    "args": {
      "External id": 842,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 639, "finished": true, "Ev Idx": 841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826160.748, "dur": 0.282,
    "args": {
      "External id": 843,"Record function id": 0, "finished": true, "Ev Idx": 842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564826163.214, "dur": 1.695,
    "args": {
      "External id": 844,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 640, "finished": true, "Ev Idx": 843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826164.163, "dur": 0.333,
    "args": {
      "External id": 845,"Record function id": 0, "finished": true, "Ev Idx": 844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826166.559, "dur": 13.163,
    "args": {
      "External id": 846,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 641, "finished": true, "Ev Idx": 845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564826167.507, "dur": 9.935,
    "args": {
      "External id": 847,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 641, "finished": true, "Ev Idx": 846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564826168.544, "dur": 2.023,
    "args": {
      "External id": 848,"Record function id": 0, "finished": true, "Ev Idx": 847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564826168.893, "dur": 1.552,
    "args": {
      "External id": 849,"Record function id": 0, "finished": true, "Ev Idx": 848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564826170.947, "dur": 6.138,
    "args": {
      "External id": 850,"Record function id": 0, "finished": true, "Ev Idx": 849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564826177.838, "dur": 1.707,
    "args": {
      "External id": 851,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 642, "finished": true, "Ev Idx": 850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564826182.200, "dur": 2.283,
    "args": {
      "External id": 852,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 643, "finished": true, "Ev Idx": 851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826183.480, "dur": 0.392,
    "args": {
      "External id": 853,"Record function id": 0, "finished": true, "Ev Idx": 852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564826186.157, "dur": 54.569,
    "args": {
      "External id": 854,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 644, "finished": true, "Ev Idx": 853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564826186.939, "dur": 1.764,
    "args": {
      "External id": 855,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 644, "finished": true, "Ev Idx": 854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826187.916, "dur": 0.341,
    "args": {
      "External id": 856,"Record function id": 0, "finished": true, "Ev Idx": 855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826189.105, "dur": 4.093,
    "args": {
      "External id": 857,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 645, "finished": true, "Ev Idx": 856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564826191.014, "dur": 2.053,
    "args": {
      "External id": 858,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 645, "finished": true, "Ev Idx": 857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564826193.750, "dur": 1.305,
    "args": {
      "External id": 859,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 646, "finished": true, "Ev Idx": 858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826194.419, "dur": 0.268,
    "args": {
      "External id": 860,"Record function id": 0, "finished": true, "Ev Idx": 859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826195.366, "dur": 1.891,
    "args": {
      "External id": 861,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 647, "finished": true, "Ev Idx": 860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564826196.228, "dur": 0.937,
    "args": {
      "External id": 862,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 647, "finished": true, "Ev Idx": 861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564826197.964, "dur": 40.487,
    "args": {
      "External id": 863,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 648, "finished": true, "Ev Idx": 862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564826215.585, "dur": 1.257,
    "args": {
      "External id": 864,"Record function id": 0, "finished": true, "Ev Idx": 863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564826239.063, "dur": 1.274,
    "args": {
      "External id": 865,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 649, "finished": true, "Ev Idx": 864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564826244.914, "dur": 11.205,
    "args": {
      "External id": 866,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 650, "finished": true, "Ev Idx": 865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826259.717, "dur": 1.927,
    "args": {
      "External id": 867,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 651, "finished": true, "Ev Idx": 866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826260.633, "dur": 0.411,
    "args": {
      "External id": 868,"Record function id": 0, "finished": true, "Ev Idx": 867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826262.463, "dur": 1.490,
    "args": {
      "External id": 869,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 651, "finished": true, "Ev Idx": 868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826263.275, "dur": 0.296,
    "args": {
      "External id": 870,"Record function id": 0, "finished": true, "Ev Idx": 869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564826264.601, "dur": 1.031,
    "args": {
      "External id": 871,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 651, "finished": true, "Ev Idx": 870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826265.056, "dur": 0.253,
    "args": {
      "External id": 872,"Record function id": 0, "finished": true, "Ev Idx": 871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564826267.536, "dur": 8.702,
    "args": {
      "External id": 873,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 651, "finished": true, "Ev Idx": 872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564826282.609, "dur": 14.122,
    "args": {
      "External id": 874,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 652, "finished": true, "Ev Idx": 873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564826283.598, "dur": 0.425,
    "args": {
      "External id": 875,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 652, "finished": true, "Ev Idx": 874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564826284.669, "dur": 11.811,
    "args": {
      "External id": 876,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 652, "finished": true, "Ev Idx": 875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564826299.998, "dur": 0.195,
    "args": {
      "External id": 877,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 653, "finished": true, "Ev Idx": 876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564826312.849, "dur": 0.634,
    "args": {
      "External id": 878,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 653, "finished": true, "Ev Idx": 877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564826315.543, "dur": 42.935,
    "args": {
      "External id": 879,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 653, "finished": true, "Ev Idx": 878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564826316.073, "dur": 2.445,
    "args": {
      "External id": 880,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 653, "finished": true, "Ev Idx": 879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826317.688, "dur": 0.343,
    "args": {
      "External id": 881,"Record function id": 0, "finished": true, "Ev Idx": 880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826318.911, "dur": 3.451,
    "args": {
      "External id": 882,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 654, "finished": true, "Ev Idx": 881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564826319.313, "dur": 2.946,
    "args": {
      "External id": 883,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 654, "finished": true, "Ev Idx": 882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564826322.838, "dur": 1.580,
    "args": {
      "External id": 884,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 655, "finished": true, "Ev Idx": 883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826323.895, "dur": 0.200,
    "args": {
      "External id": 885,"Record function id": 0, "finished": true, "Ev Idx": 884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826324.750, "dur": 1.185,
    "args": {
      "External id": 886,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 656, "finished": true, "Ev Idx": 885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564826325.046, "dur": 0.817,
    "args": {
      "External id": 887,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 656, "finished": true, "Ev Idx": 886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564826326.308, "dur": 30.138,
    "args": {
      "External id": 888,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 657, "finished": true, "Ev Idx": 887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564826337.547, "dur": 1.184,
    "args": {
      "External id": 889,"Record function id": 0, "finished": true, "Ev Idx": 888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564826357.030, "dur": 1.142,
    "args": {
      "External id": 890,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 658, "finished": true, "Ev Idx": 889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564826360.633, "dur": 1.739,
    "args": {
      "External id": 891,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 659, "finished": true, "Ev Idx": 890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826361.495, "dur": 0.344,
    "args": {
      "External id": 892,"Record function id": 0, "finished": true, "Ev Idx": 891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564826364.835, "dur": 12.263,
    "args": {
      "External id": 893,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 660, "finished": true, "Ev Idx": 892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564826365.639, "dur": 11.315,
    "args": {
      "External id": 894,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 660, "finished": true, "Ev Idx": 893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564826366.898, "dur": 2.079,
    "args": {
      "External id": 895,"Record function id": 0, "finished": true, "Ev Idx": 894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564826367.516, "dur": 1.330,
    "args": {
      "External id": 896,"Record function id": 0, "finished": true, "Ev Idx": 895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564826369.358, "dur": 7.125,
    "args": {
      "External id": 897,"Record function id": 0, "finished": true, "Ev Idx": 896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826383.017, "dur": 1.961,
    "args": {
      "External id": 898,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 661, "finished": true, "Ev Idx": 897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564826383.586, "dur": 1.286,
    "args": {
      "External id": 899,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 661, "finished": true, "Ev Idx": 898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564826401.156, "dur": 36.719,
    "args": {
      "External id": 900,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 662, "finished": true, "Ev Idx": 899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564826401.882, "dur": 3.286,
    "args": {
      "External id": 901,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 662, "finished": true, "Ev Idx": 900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564826403.598, "dur": 1.037,
    "args": {
      "External id": 902,"Record function id": 0, "finished": true, "Ev Idx": 901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826404.037, "dur": 0.503,
    "args": {
      "External id": 903,"Record function id": 0, "finished": true, "Ev Idx": 902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564826405.554, "dur": 32.094,
    "args": {
      "External id": 904,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 663, "finished": true, "Ev Idx": 903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826406.257, "dur": 1.452,
    "args": {
      "External id": 905,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 663, "finished": true, "Ev Idx": 904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564826406.537, "dur": 1.076,
    "args": {
      "External id": 906,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 663, "finished": true, "Ev Idx": 905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564826408.320, "dur": 27.538,
    "args": {
      "External id": 907,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 664, "finished": true, "Ev Idx": 906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564826419.887, "dur": 1.088,
    "args": {
      "External id": 908,"Record function id": 0, "finished": true, "Ev Idx": 907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564826436.390, "dur": 0.968,
    "args": {
      "External id": 909,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 665, "finished": true, "Ev Idx": 908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564826443.193, "dur": 8.976,
    "args": {
      "External id": 910,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 666, "finished": true, "Ev Idx": 909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564826461.672, "dur": 0.235,
    "args": {
      "External id": 911,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 667, "finished": true, "Ev Idx": 910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564826464.220, "dur": 12.472,
    "args": {
      "External id": 912,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 667, "finished": true, "Ev Idx": 911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564826466.158, "dur": 0.423,
    "args": {
      "External id": 913,"Record function id": 0, "finished": true, "Ev Idx": 912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564826467.006, "dur": 0.106,
    "args": {
      "External id": 914,"Record function id": 0, "finished": true, "Ev Idx": 913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564826479.471, "dur": 13.251,
    "args": {
      "External id": 915,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 668, "finished": true, "Ev Idx": 914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564826496.430, "dur": 10.318,
    "args": {
      "External id": 916,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 669, "finished": true, "Ev Idx": 915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564826508.717, "dur": 9.793,
    "args": {
      "External id": 917,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 670, "finished": true, "Ev Idx": 916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564826520.241, "dur": 8.347,
    "args": {
      "External id": 918,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 671, "finished": true, "Ev Idx": 917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564826531.063, "dur": 0.114,
    "args": {
      "External id": 919,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 672, "finished": true, "Ev Idx": 918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564826532.193, "dur": 9.750,
    "args": {
      "External id": 920,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 672, "finished": true, "Ev Idx": 919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564826555.691, "dur": 47.302,
    "args": {
      "External id": 921,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 673, "finished": true, "Ev Idx": 920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564826556.333, "dur": 3.436,
    "args": {
      "External id": 922,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 673, "finished": true, "Ev Idx": 921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564826557.377, "dur": 1.772,
    "args": {
      "External id": 923,"Record function id": 0, "finished": true, "Ev Idx": 922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826558.670, "dur": 0.377,
    "args": {
      "External id": 924,"Record function id": 0, "finished": true, "Ev Idx": 923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564826560.108, "dur": 42.753,
    "args": {
      "External id": 925,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 674, "finished": true, "Ev Idx": 924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826560.634, "dur": 2.761,
    "args": {
      "External id": 926,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 674, "finished": true, "Ev Idx": 925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564826560.946, "dur": 2.332,
    "args": {
      "External id": 927,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 674, "finished": true, "Ev Idx": 926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564826563.708, "dur": 37.231,
    "args": {
      "External id": 928,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 675, "finished": true, "Ev Idx": 927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564826573.783, "dur": 1.217,
    "args": {
      "External id": 929,"Record function id": 0, "finished": true, "Ev Idx": 928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564826601.607, "dur": 1.010,
    "args": {
      "External id": 930,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 676, "finished": true, "Ev Idx": 929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564826611.813, "dur": 67.242,
    "args": {
      "External id": 931,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 677, "finished": true, "Ev Idx": 930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564826688.039, "dur": 46.541,
    "args": {
      "External id": 932,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 678, "finished": true, "Ev Idx": 931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564826688.529, "dur": 4.123,
    "args": {
      "External id": 933,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 678, "finished": true, "Ev Idx": 932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564826689.999, "dur": 1.559,
    "args": {
      "External id": 934,"Record function id": 0, "finished": true, "Ev Idx": 933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826690.540, "dur": 0.865,
    "args": {
      "External id": 935,"Record function id": 0, "finished": true, "Ev Idx": 934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564826693.018, "dur": 41.434,
    "args": {
      "External id": 936,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 679, "finished": true, "Ev Idx": 935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826693.677, "dur": 2.520,
    "args": {
      "External id": 937,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 679, "finished": true, "Ev Idx": 936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564826694.108, "dur": 1.969,
    "args": {
      "External id": 938,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 679, "finished": true, "Ev Idx": 937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564826696.861, "dur": 35.473,
    "args": {
      "External id": 939,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 680, "finished": true, "Ev Idx": 938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564826709.317, "dur": 1.200,
    "args": {
      "External id": 940,"Record function id": 0, "finished": true, "Ev Idx": 939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564826732.944, "dur": 1.200,
    "args": {
      "External id": 941,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 681, "finished": true, "Ev Idx": 940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564826736.544, "dur": 38.438,
    "args": {
      "External id": 942,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 682, "finished": true, "Ev Idx": 941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564826782.887, "dur": 43.380,
    "args": {
      "External id": 943,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 683, "finished": true, "Ev Idx": 942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564826783.441, "dur": 4.352,
    "args": {
      "External id": 944,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 683, "finished": true, "Ev Idx": 943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564826785.232, "dur": 1.575,
    "args": {
      "External id": 945,"Record function id": 0, "finished": true, "Ev Idx": 944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826785.936, "dur": 0.689,
    "args": {
      "External id": 946,"Record function id": 0, "finished": true, "Ev Idx": 945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564826788.286, "dur": 37.850,
    "args": {
      "External id": 947,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 684, "finished": true, "Ev Idx": 946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826788.954, "dur": 1.718,
    "args": {
      "External id": 948,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 684, "finished": true, "Ev Idx": 947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564826789.434, "dur": 1.127,
    "args": {
      "External id": 949,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 684, "finished": true, "Ev Idx": 948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564826791.105, "dur": 33.068,
    "args": {
      "External id": 950,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 685, "finished": true, "Ev Idx": 949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564826800.881, "dur": 1.057,
    "args": {
      "External id": 951,"Record function id": 0, "finished": true, "Ev Idx": 950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564826824.694, "dur": 1.142,
    "args": {
      "External id": 952,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 686, "finished": true, "Ev Idx": 951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564826829.148, "dur": 8.028,
    "args": {
      "External id": 953,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 687, "finished": true, "Ev Idx": 952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564826857.679, "dur": 0.370,
    "args": {
      "External id": 954,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 688, "finished": true, "Ev Idx": 953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564826860.014, "dur": 9.891,
    "args": {
      "External id": 955,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 688, "finished": true, "Ev Idx": 954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564826861.468, "dur": 0.512,
    "args": {
      "External id": 956,"Record function id": 0, "finished": true, "Ev Idx": 955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564826862.294, "dur": 0.215,
    "args": {
      "External id": 957,"Record function id": 0, "finished": true, "Ev Idx": 956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564826872.622, "dur": 12.772,
    "args": {
      "External id": 958,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 689, "finished": true, "Ev Idx": 957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564826889.280, "dur": 8.748,
    "args": {
      "External id": 959,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 690, "finished": true, "Ev Idx": 958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564826899.939, "dur": 9.143,
    "args": {
      "External id": 960,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 691, "finished": true, "Ev Idx": 959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564826910.782, "dur": 8.773,
    "args": {
      "External id": 961,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 692, "finished": true, "Ev Idx": 960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564826921.883, "dur": 0.281,
    "args": {
      "External id": 962,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 693, "finished": true, "Ev Idx": 961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564826922.935, "dur": 7.858,
    "args": {
      "External id": 963,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 693, "finished": true, "Ev Idx": 962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564826949.010, "dur": 38.470,
    "args": {
      "External id": 964,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 694, "finished": true, "Ev Idx": 963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564826949.550, "dur": 4.730,
    "args": {
      "External id": 965,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 694, "finished": true, "Ev Idx": 964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564826951.924, "dur": 1.247,
    "args": {
      "External id": 966,"Record function id": 0, "finished": true, "Ev Idx": 965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826952.519, "dur": 0.523,
    "args": {
      "External id": 967,"Record function id": 0, "finished": true, "Ev Idx": 966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564826954.741, "dur": 32.610,
    "args": {
      "External id": 968,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 695, "finished": true, "Ev Idx": 967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564826955.245, "dur": 1.736,
    "args": {
      "External id": 969,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 695, "finished": true, "Ev Idx": 968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564826955.711, "dur": 1.139,
    "args": {
      "External id": 970,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 695, "finished": true, "Ev Idx": 969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564826957.309, "dur": 28.007,
    "args": {
      "External id": 971,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 696, "finished": true, "Ev Idx": 970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564826968.977, "dur": 1.118,
    "args": {
      "External id": 972,"Record function id": 0, "finished": true, "Ev Idx": 971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564826985.928, "dur": 1.150,
    "args": {
      "External id": 973,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 697, "finished": true, "Ev Idx": 972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564826991.017, "dur": 1.437,
    "args": {
      "External id": 974,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 698, "finished": true, "Ev Idx": 973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564826994.575, "dur": 2.557,
    "args": {
      "External id": 975,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 699, "finished": true, "Ev Idx": 974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564826995.783, "dur": 0.483,
    "args": {
      "External id": 976,"Record function id": 0, "finished": true, "Ev Idx": 975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564827004.127, "dur": 38.833,
    "args": {
      "External id": 977,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 700, "finished": true, "Ev Idx": 976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564827004.432, "dur": 2.315,
    "args": {
      "External id": 978,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 700, "finished": true, "Ev Idx": 977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564827005.392, "dur": 0.880,
    "args": {
      "External id": 979,"Record function id": 0, "finished": true, "Ev Idx": 978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827005.867, "dur": 0.283,
    "args": {
      "External id": 980,"Record function id": 0, "finished": true, "Ev Idx": 979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564827007.072, "dur": 35.789,
    "args": {
      "External id": 981,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 701, "finished": true, "Ev Idx": 980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827007.466, "dur": 1.191,
    "args": {
      "External id": 982,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 701, "finished": true, "Ev Idx": 981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827007.742, "dur": 0.823,
    "args": {
      "External id": 983,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 701, "finished": true, "Ev Idx": 982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564827008.960, "dur": 31.902,
    "args": {
      "External id": 984,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 702, "finished": true, "Ev Idx": 983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564827018.781, "dur": 1.179,
    "args": {
      "External id": 985,"Record function id": 0, "finished": true, "Ev Idx": 984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564827041.418, "dur": 1.166,
    "args": {
      "External id": 986,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 703, "finished": true, "Ev Idx": 985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827045.162, "dur": 1.204,
    "args": {
      "External id": 987,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 704, "finished": true, "Ev Idx": 986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564827047.853, "dur": 2.056,
    "args": {
      "External id": 988,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 705, "finished": true, "Ev Idx": 987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827048.814, "dur": 0.472,
    "args": {
      "External id": 989,"Record function id": 0, "finished": true, "Ev Idx": 988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564827056.758, "dur": 38.251,
    "args": {
      "External id": 990,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 706, "finished": true, "Ev Idx": 989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564827057.206, "dur": 1.998,
    "args": {
      "External id": 991,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 706, "finished": true, "Ev Idx": 990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564827057.955, "dur": 0.839,
    "args": {
      "External id": 992,"Record function id": 0, "finished": true, "Ev Idx": 991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827058.424, "dur": 0.279,
    "args": {
      "External id": 993,"Record function id": 0, "finished": true, "Ev Idx": 992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564827059.510, "dur": 35.403,
    "args": {
      "External id": 994,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 707, "finished": true, "Ev Idx": 993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827060.059, "dur": 1.326,
    "args": {
      "External id": 995,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 707, "finished": true, "Ev Idx": 994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827060.468, "dur": 0.819,
    "args": {
      "External id": 996,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 707, "finished": true, "Ev Idx": 995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564827061.700, "dur": 31.485,
    "args": {
      "External id": 997,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 708, "finished": true, "Ev Idx": 996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564827074.030, "dur": 1.034,
    "args": {
      "External id": 998,"Record function id": 0, "finished": true, "Ev Idx": 997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564827093.710, "dur": 0.929,
    "args": {
      "External id": 999,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 709, "finished": true, "Ev Idx": 998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827097.199, "dur": 1.138,
    "args": {
      "External id": 1000,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 710, "finished": true, "Ev Idx": 999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564827099.736, "dur": 1.937,
    "args": {
      "External id": 1001,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 711, "finished": true, "Ev Idx": 1000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827100.834, "dur": 0.351,
    "args": {
      "External id": 1002,"Record function id": 0, "finished": true, "Ev Idx": 1001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564827104.700, "dur": 2.187,
    "args": {
      "External id": 1003,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 712, "finished": true, "Ev Idx": 1002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827106.111, "dur": 0.304,
    "args": {
      "External id": 1004,"Record function id": 0, "finished": true, "Ev Idx": 1003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564827108.056, "dur": 1.055,
    "args": {
      "External id": 1005,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 712, "finished": true, "Ev Idx": 1004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827108.519, "dur": 0.294,
    "args": {
      "External id": 1006,"Record function id": 0, "finished": true, "Ev Idx": 1005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564827109.916, "dur": 10.128,
    "args": {
      "External id": 1007,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 712, "finished": true, "Ev Idx": 1006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827125.352, "dur": 5.344,
    "args": {
      "External id": 1008,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 713, "finished": true, "Ev Idx": 1007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827127.760, "dur": 0.960,
    "args": {
      "External id": 1009,"Record function id": 0, "finished": true, "Ev Idx": 1008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827132.991, "dur": 2.391,
    "args": {
      "External id": 1010,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 714, "finished": true, "Ev Idx": 1009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827134.427, "dur": 0.351,
    "args": {
      "External id": 1011,"Record function id": 0, "finished": true, "Ev Idx": 1010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564827137.319, "dur": 9.668,
    "args": {
      "External id": 1012,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 715, "finished": true, "Ev Idx": 1011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564827152.728, "dur": 17.426,
    "args": {
      "External id": 1013,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 716, "finished": true, "Ev Idx": 1012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564827173.211, "dur": 7.940,
    "args": {
      "External id": 1014,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 717, "finished": true, "Ev Idx": 1013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564827182.765, "dur": 9.794,
    "args": {
      "External id": 1015,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 718, "finished": true, "Ev Idx": 1014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564827194.798, "dur": 6.740,
    "args": {
      "External id": 1016,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 719, "finished": true, "Ev Idx": 1015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827203.914, "dur": 2.309,
    "args": {
      "External id": 1017,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 720, "finished": true, "Ev Idx": 1016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827205.074, "dur": 0.502,
    "args": {
      "External id": 1018,"Record function id": 0, "finished": true, "Ev Idx": 1017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827207.754, "dur": 1.825,
    "args": {
      "External id": 1019,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 721, "finished": true, "Ev Idx": 1018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827208.732, "dur": 0.336,
    "args": {
      "External id": 1020,"Record function id": 0, "finished": true, "Ev Idx": 1019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564827210.630, "dur": 8.198,
    "args": {
      "External id": 1021,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 722, "finished": true, "Ev Idx": 1020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564827220.791, "dur": 9.297,
    "args": {
      "External id": 1022,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 723, "finished": true, "Ev Idx": 1021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564827232.362, "dur": 7.996,
    "args": {
      "External id": 1023,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 724, "finished": true, "Ev Idx": 1022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564827241.927, "dur": 6.738,
    "args": {
      "External id": 1024,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 725, "finished": true, "Ev Idx": 1023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827305.474, "dur": 2.502,
    "args": {
      "External id": 1025,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 726, "finished": true, "Ev Idx": 1024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827306.696, "dur": 0.529,
    "args": {
      "External id": 1026,"Record function id": 0, "finished": true, "Ev Idx": 1025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827308.667, "dur": 1.601,
    "args": {
      "External id": 1027,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 727, "finished": true, "Ev Idx": 1026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827309.389, "dur": 0.358,
    "args": {
      "External id": 1028,"Record function id": 0, "finished": true, "Ev Idx": 1027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564827311.031, "dur": 2.288,
    "args": {
      "External id": 1029,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 728, "finished": true, "Ev Idx": 1028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827311.941, "dur": 0.690,
    "args": {
      "External id": 1030,"Record function id": 0, "finished": true, "Ev Idx": 1029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827314.003, "dur": 2.868,
    "args": {
      "External id": 1031,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 729, "finished": true, "Ev Idx": 1030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827315.958, "dur": 0.387,
    "args": {
      "External id": 1032,"Record function id": 0, "finished": true, "Ev Idx": 1031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827317.492, "dur": 2.476,
    "args": {
      "External id": 1033,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 730, "finished": true, "Ev Idx": 1032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827319.148, "dur": 0.320,
    "args": {
      "External id": 1034,"Record function id": 0, "finished": true, "Ev Idx": 1033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564827322.941, "dur": 2.869,
    "args": {
      "External id": 1035,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 731, "finished": true, "Ev Idx": 1034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827324.735, "dur": 0.391,
    "args": {
      "External id": 1036,"Record function id": 0, "finished": true, "Ev Idx": 1035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827328.207, "dur": 19.200,
    "args": {
      "External id": 1037,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 732, "finished": true, "Ev Idx": 1036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564827330.604, "dur": 14.606,
    "args": {
      "External id": 1038,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 732, "finished": true, "Ev Idx": 1037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564827332.111, "dur": 2.623,
    "args": {
      "External id": 1039,"Record function id": 0, "finished": true, "Ev Idx": 1038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564827332.836, "dur": 1.668,
    "args": {
      "External id": 1040,"Record function id": 0, "finished": true, "Ev Idx": 1039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564827335.360, "dur": 9.228,
    "args": {
      "External id": 1041,"Record function id": 0, "finished": true, "Ev Idx": 1040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564827345.822, "dur": 1.300,
    "args": {
      "External id": 1042,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 733, "finished": true, "Ev Idx": 1041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827350.676, "dur": 2.564,
    "args": {
      "External id": 1043,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 734, "finished": true, "Ev Idx": 1042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827352.226, "dur": 0.439,
    "args": {
      "External id": 1044,"Record function id": 0, "finished": true, "Ev Idx": 1043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827353.846, "dur": 2.038,
    "args": {
      "External id": 1045,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 735, "finished": true, "Ev Idx": 1044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827355.103, "dur": 0.300,
    "args": {
      "External id": 1046,"Record function id": 0, "finished": true, "Ev Idx": 1045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564827356.430, "dur": 1.457,
    "args": {
      "External id": 1047,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 736, "finished": true, "Ev Idx": 1046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827357.125, "dur": 0.308,
    "args": {
      "External id": 1048,"Record function id": 0, "finished": true, "Ev Idx": 1047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827358.514, "dur": 1.479,
    "args": {
      "External id": 1049,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 737, "finished": true, "Ev Idx": 1048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827359.225, "dur": 0.310,
    "args": {
      "External id": 1050,"Record function id": 0, "finished": true, "Ev Idx": 1049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827360.596, "dur": 1.380,
    "args": {
      "External id": 1051,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 738, "finished": true, "Ev Idx": 1050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827361.247, "dur": 0.277,
    "args": {
      "External id": 1052,"Record function id": 0, "finished": true, "Ev Idx": 1051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564827363.552, "dur": 1.513,
    "args": {
      "External id": 1053,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 739, "finished": true, "Ev Idx": 1052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827364.346, "dur": 0.316,
    "args": {
      "External id": 1054,"Record function id": 0, "finished": true, "Ev Idx": 1053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827366.541, "dur": 12.251,
    "args": {
      "External id": 1055,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 740, "finished": true, "Ev Idx": 1054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564827367.487, "dur": 9.828,
    "args": {
      "External id": 1056,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 740, "finished": true, "Ev Idx": 1055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564827369.076, "dur": 1.848,
    "args": {
      "External id": 1057,"Record function id": 0, "finished": true, "Ev Idx": 1056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564827369.610, "dur": 1.208,
    "args": {
      "External id": 1058,"Record function id": 0, "finished": true, "Ev Idx": 1057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564827371.253, "dur": 5.739,
    "args": {
      "External id": 1059,"Record function id": 0, "finished": true, "Ev Idx": 1058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564827377.705, "dur": 0.877,
    "args": {
      "External id": 1060,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 741, "finished": true, "Ev Idx": 1059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564827381.047, "dur": 1.885,
    "args": {
      "External id": 1061,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 742, "finished": true, "Ev Idx": 1060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827382.042, "dur": 0.393,
    "args": {
      "External id": 1062,"Record function id": 0, "finished": true, "Ev Idx": 1061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564827384.523, "dur": 47.410,
    "args": {
      "External id": 1063,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 743, "finished": true, "Ev Idx": 1062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564827385.190, "dur": 1.456,
    "args": {
      "External id": 1064,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 743, "finished": true, "Ev Idx": 1063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827385.940, "dur": 0.305,
    "args": {
      "External id": 1065,"Record function id": 0, "finished": true, "Ev Idx": 1064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827387.010, "dur": 3.653,
    "args": {
      "External id": 1066,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 744, "finished": true, "Ev Idx": 1065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564827388.684, "dur": 1.863,
    "args": {
      "External id": 1067,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 744, "finished": true, "Ev Idx": 1066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564827391.134, "dur": 1.748,
    "args": {
      "External id": 1068,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 745, "finished": true, "Ev Idx": 1067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827392.213, "dur": 0.282,
    "args": {
      "External id": 1069,"Record function id": 0, "finished": true, "Ev Idx": 1068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827393.197, "dur": 2.709,
    "args": {
      "External id": 1070,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 746, "finished": true, "Ev Idx": 1069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564827394.182, "dur": 1.635,
    "args": {
      "External id": 1071,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 746, "finished": true, "Ev Idx": 1070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564827396.535, "dur": 33.335,
    "args": {
      "External id": 1072,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 747, "finished": true, "Ev Idx": 1071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564827410.052, "dur": 1.210,
    "args": {
      "External id": 1073,"Record function id": 0, "finished": true, "Ev Idx": 1072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564827430.485, "dur": 1.080,
    "args": {
      "External id": 1074,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 748, "finished": true, "Ev Idx": 1073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564827435.448, "dur": 9.566,
    "args": {
      "External id": 1075,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 749, "finished": true, "Ev Idx": 1074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827448.200, "dur": 1.785,
    "args": {
      "External id": 1076,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 750, "finished": true, "Ev Idx": 1075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827448.974, "dur": 0.423,
    "args": {
      "External id": 1077,"Record function id": 0, "finished": true, "Ev Idx": 1076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827450.593, "dur": 1.303,
    "args": {
      "External id": 1078,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 750, "finished": true, "Ev Idx": 1077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827451.208, "dur": 0.291,
    "args": {
      "External id": 1079,"Record function id": 0, "finished": true, "Ev Idx": 1078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564827452.550, "dur": 0.984,
    "args": {
      "External id": 1080,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 750, "finished": true, "Ev Idx": 1079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827453.004, "dur": 0.215,
    "args": {
      "External id": 1081,"Record function id": 0, "finished": true, "Ev Idx": 1080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564827455.197, "dur": 8.181,
    "args": {
      "External id": 1082,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 750, "finished": true, "Ev Idx": 1081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564827468.799, "dur": 12.890,
    "args": {
      "External id": 1083,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 751, "finished": true, "Ev Idx": 1082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564827469.700, "dur": 0.394,
    "args": {
      "External id": 1084,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 751, "finished": true, "Ev Idx": 1083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564827470.728, "dur": 10.681,
    "args": {
      "External id": 1085,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 751, "finished": true, "Ev Idx": 1084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564827484.547, "dur": 0.164,
    "args": {
      "External id": 1086,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 752, "finished": true, "Ev Idx": 1085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564827495.634, "dur": 0.477,
    "args": {
      "External id": 1087,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 752, "finished": true, "Ev Idx": 1086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564827497.843, "dur": 42.880,
    "args": {
      "External id": 1088,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 752, "finished": true, "Ev Idx": 1087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564827498.509, "dur": 2.967,
    "args": {
      "External id": 1089,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 752, "finished": true, "Ev Idx": 1088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827500.634, "dur": 0.392,
    "args": {
      "External id": 1090,"Record function id": 0, "finished": true, "Ev Idx": 1089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827501.984, "dur": 2.579,
    "args": {
      "External id": 1091,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 753, "finished": true, "Ev Idx": 1090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827502.372, "dur": 2.065,
    "args": {
      "External id": 1092,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 753, "finished": true, "Ev Idx": 1091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564827505.017, "dur": 1.204,
    "args": {
      "External id": 1093,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 754, "finished": true, "Ev Idx": 1092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827505.606, "dur": 0.255,
    "args": {
      "External id": 1094,"Record function id": 0, "finished": true, "Ev Idx": 1093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827506.498, "dur": 1.130,
    "args": {
      "External id": 1095,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 755, "finished": true, "Ev Idx": 1094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827506.743, "dur": 0.806,
    "args": {
      "External id": 1096,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 755, "finished": true, "Ev Idx": 1095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564827508.032, "dur": 30.746,
    "args": {
      "External id": 1097,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 756, "finished": true, "Ev Idx": 1096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564827521.133, "dur": 1.154,
    "args": {
      "External id": 1098,"Record function id": 0, "finished": true, "Ev Idx": 1097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564827539.344, "dur": 1.081,
    "args": {
      "External id": 1099,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 757, "finished": true, "Ev Idx": 1098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564827542.772, "dur": 1.822,
    "args": {
      "External id": 1100,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 758, "finished": true, "Ev Idx": 1099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827543.720, "dur": 0.391,
    "args": {
      "External id": 1101,"Record function id": 0, "finished": true, "Ev Idx": 1100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564827547.024, "dur": 11.230,
    "args": {
      "External id": 1102,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 759, "finished": true, "Ev Idx": 1101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564827547.553, "dur": 10.567,
    "args": {
      "External id": 1103,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 759, "finished": true, "Ev Idx": 1102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564827548.606, "dur": 2.078,
    "args": {
      "External id": 1104,"Record function id": 0, "finished": true, "Ev Idx": 1103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564827548.996, "dur": 1.544,
    "args": {
      "External id": 1105,"Record function id": 0, "finished": true, "Ev Idx": 1104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564827551.028, "dur": 6.694,
    "args": {
      "External id": 1106,"Record function id": 0, "finished": true, "Ev Idx": 1105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827564.187, "dur": 1.736,
    "args": {
      "External id": 1107,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 760, "finished": true, "Ev Idx": 1106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827564.506, "dur": 1.303,
    "args": {
      "External id": 1108,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 760, "finished": true, "Ev Idx": 1107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564827580.829, "dur": 36.228,
    "args": {
      "External id": 1109,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 761, "finished": true, "Ev Idx": 1108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564827581.598, "dur": 3.066,
    "args": {
      "External id": 1110,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 761, "finished": true, "Ev Idx": 1109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564827583.074, "dur": 0.987,
    "args": {
      "External id": 1111,"Record function id": 0, "finished": true, "Ev Idx": 1110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827583.567, "dur": 0.339,
    "args": {
      "External id": 1112,"Record function id": 0, "finished": true, "Ev Idx": 1111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564827585.031, "dur": 31.840,
    "args": {
      "External id": 1113,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 762, "finished": true, "Ev Idx": 1112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827585.498, "dur": 1.936,
    "args": {
      "External id": 1114,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 762, "finished": true, "Ev Idx": 1113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827585.748, "dur": 1.585,
    "args": {
      "External id": 1115,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 762, "finished": true, "Ev Idx": 1114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564827587.832, "dur": 26.820,
    "args": {
      "External id": 1116,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 763, "finished": true, "Ev Idx": 1115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564827599.712, "dur": 1.088,
    "args": {
      "External id": 1117,"Record function id": 0, "finished": true, "Ev Idx": 1116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564827615.295, "dur": 1.191,
    "args": {
      "External id": 1118,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 764, "finished": true, "Ev Idx": 1117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564827622.045, "dur": 8.533,
    "args": {
      "External id": 1119,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 765, "finished": true, "Ev Idx": 1118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564827639.888, "dur": 0.169,
    "args": {
      "External id": 1120,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 766, "finished": true, "Ev Idx": 1119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564827642.225, "dur": 11.632,
    "args": {
      "External id": 1121,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 766, "finished": true, "Ev Idx": 1120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564827644.032, "dur": 0.479,
    "args": {
      "External id": 1122,"Record function id": 0, "finished": true, "Ev Idx": 1121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564827644.930, "dur": 0.061,
    "args": {
      "External id": 1123,"Record function id": 0, "finished": true, "Ev Idx": 1122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564827656.717, "dur": 13.225,
    "args": {
      "External id": 1124,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 767, "finished": true, "Ev Idx": 1123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564827673.498, "dur": 9.233,
    "args": {
      "External id": 1125,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 768, "finished": true, "Ev Idx": 1124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564827684.847, "dur": 9.218,
    "args": {
      "External id": 1126,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 769, "finished": true, "Ev Idx": 1125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564827695.910, "dur": 9.914,
    "args": {
      "External id": 1127,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 770, "finished": true, "Ev Idx": 1126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564827708.512, "dur": 0.094,
    "args": {
      "External id": 1128,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 771, "finished": true, "Ev Idx": 1127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564827709.465, "dur": 73.717,
    "args": {
      "External id": 1129,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 771, "finished": true, "Ev Idx": 1128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564827804.998, "dur": 55.227,
    "args": {
      "External id": 1130,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 772, "finished": true, "Ev Idx": 1129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564827805.747, "dur": 3.883,
    "args": {
      "External id": 1131,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 772, "finished": true, "Ev Idx": 1130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564827806.910, "dur": 1.925,
    "args": {
      "External id": 1132,"Record function id": 0, "finished": true, "Ev Idx": 1131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827807.854, "dur": 0.755,
    "args": {
      "External id": 1133,"Record function id": 0, "finished": true, "Ev Idx": 1132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564827810.194, "dur": 49.894,
    "args": {
      "External id": 1134,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 773, "finished": true, "Ev Idx": 1133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827810.796, "dur": 2.037,
    "args": {
      "External id": 1135,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 773, "finished": true, "Ev Idx": 1134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827811.265, "dur": 1.452,
    "args": {
      "External id": 1136,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 773, "finished": true, "Ev Idx": 1135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564827813.238, "dur": 44.674,
    "args": {
      "External id": 1137,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 774, "finished": true, "Ev Idx": 1136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564827827.015, "dur": 1.428,
    "args": {
      "External id": 1138,"Record function id": 0, "finished": true, "Ev Idx": 1137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564827858.593, "dur": 1.120,
    "args": {
      "External id": 1139,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 775, "finished": true, "Ev Idx": 1138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564827869.376, "dur": 10.721,
    "args": {
      "External id": 1140,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 776, "finished": true, "Ev Idx": 1139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564827888.246, "dur": 80.588,
    "args": {
      "External id": 1141,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 777, "finished": true, "Ev Idx": 1140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564827888.568, "dur": 2.780,
    "args": {
      "External id": 1142,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 777, "finished": true, "Ev Idx": 1141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564827889.479, "dur": 1.364,
    "args": {
      "External id": 1143,"Record function id": 0, "finished": true, "Ev Idx": 1142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827890.248, "dur": 0.465,
    "args": {
      "External id": 1144,"Record function id": 0, "finished": true, "Ev Idx": 1143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564827891.899, "dur": 76.808,
    "args": {
      "External id": 1145,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 778, "finished": true, "Ev Idx": 1144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827892.310, "dur": 1.983,
    "args": {
      "External id": 1146,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 778, "finished": true, "Ev Idx": 1145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827892.751, "dur": 1.418,
    "args": {
      "External id": 1147,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 778, "finished": true, "Ev Idx": 1146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564827894.659, "dur": 71.852,
    "args": {
      "External id": 1148,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 779, "finished": true, "Ev Idx": 1147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564827941.271, "dur": 1.300,
    "args": {
      "External id": 1149,"Record function id": 0, "finished": true, "Ev Idx": 1148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564827967.186, "dur": 1.185,
    "args": {
      "External id": 1150,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 780, "finished": true, "Ev Idx": 1149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564827971.042, "dur": 8.853,
    "args": {
      "External id": 1151,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 781, "finished": true, "Ev Idx": 1150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564827987.062, "dur": 38.966,
    "args": {
      "External id": 1152,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 782, "finished": true, "Ev Idx": 1151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564827987.568, "dur": 2.744,
    "args": {
      "External id": 1153,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 782, "finished": true, "Ev Idx": 1152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564827988.462, "dur": 1.268,
    "args": {
      "External id": 1154,"Record function id": 0, "finished": true, "Ev Idx": 1153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564827989.176, "dur": 0.427,
    "args": {
      "External id": 1155,"Record function id": 0, "finished": true, "Ev Idx": 1154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564827990.652, "dur": 35.281,
    "args": {
      "External id": 1156,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 783, "finished": true, "Ev Idx": 1155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564827991.303, "dur": 1.762,
    "args": {
      "External id": 1157,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 783, "finished": true, "Ev Idx": 1156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564827991.593, "dur": 1.365,
    "args": {
      "External id": 1158,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 783, "finished": true, "Ev Idx": 1157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564827993.387, "dur": 30.678,
    "args": {
      "External id": 1159,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 784, "finished": true, "Ev Idx": 1158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564828003.199, "dur": 1.174,
    "args": {
      "External id": 1160,"Record function id": 0, "finished": true, "Ev Idx": 1159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564828024.696, "dur": 0.973,
    "args": {
      "External id": 1161,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 785, "finished": true, "Ev Idx": 1160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564828028.800, "dur": 8.644,
    "args": {
      "External id": 1162,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 786, "finished": true, "Ev Idx": 1161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564828058.658, "dur": 0.267,
    "args": {
      "External id": 1163,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 787, "finished": true, "Ev Idx": 1162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564828061.049, "dur": 10.902,
    "args": {
      "External id": 1164,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 787, "finished": true, "Ev Idx": 1163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564828062.500, "dur": 0.420,
    "args": {
      "External id": 1165,"Record function id": 0, "finished": true, "Ev Idx": 1164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564828063.319, "dur": 0.074,
    "args": {
      "External id": 1166,"Record function id": 0, "finished": true, "Ev Idx": 1165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564828074.614, "dur": 13.855,
    "args": {
      "External id": 1167,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 788, "finished": true, "Ev Idx": 1166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564828092.280, "dur": 8.802,
    "args": {
      "External id": 1168,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 789, "finished": true, "Ev Idx": 1167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564828102.739, "dur": 8.061,
    "args": {
      "External id": 1169,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 790, "finished": true, "Ev Idx": 1168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564828112.528, "dur": 9.309,
    "args": {
      "External id": 1170,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 791, "finished": true, "Ev Idx": 1169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564828124.275, "dur": 0.231,
    "args": {
      "External id": 1171,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 792, "finished": true, "Ev Idx": 1170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564828125.290, "dur": 8.762,
    "args": {
      "External id": 1172,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 792, "finished": true, "Ev Idx": 1171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564828153.353, "dur": 80.392,
    "args": {
      "External id": 1173,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 793, "finished": true, "Ev Idx": 1172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564828154.270, "dur": 4.454,
    "args": {
      "External id": 1174,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 793, "finished": true, "Ev Idx": 1173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564828156.638, "dur": 1.477,
    "args": {
      "External id": 1175,"Record function id": 0, "finished": true, "Ev Idx": 1174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828157.498, "dur": 0.468,
    "args": {
      "External id": 1176,"Record function id": 0, "finished": true, "Ev Idx": 1175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564828159.040, "dur": 74.617,
    "args": {
      "External id": 1177,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 794, "finished": true, "Ev Idx": 1176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828159.589, "dur": 1.749,
    "args": {
      "External id": 1178,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 794, "finished": true, "Ev Idx": 1177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828160.022, "dur": 1.218,
    "args": {
      "External id": 1179,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 794, "finished": true, "Ev Idx": 1178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564828161.644, "dur": 69.620,
    "args": {
      "External id": 1180,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 795, "finished": true, "Ev Idx": 1179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564828212.111, "dur": 1.282,
    "args": {
      "External id": 1181,"Record function id": 0, "finished": true, "Ev Idx": 1180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564828231.987, "dur": 1.326,
    "args": {
      "External id": 1182,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 796, "finished": true, "Ev Idx": 1181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828237.634, "dur": 1.623,
    "args": {
      "External id": 1183,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 797, "finished": true, "Ev Idx": 1182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564828241.226, "dur": 2.907,
    "args": {
      "External id": 1184,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 798, "finished": true, "Ev Idx": 1183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828242.511, "dur": 0.644,
    "args": {
      "External id": 1185,"Record function id": 0, "finished": true, "Ev Idx": 1184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564828252.477, "dur": 44.698,
    "args": {
      "External id": 1186,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 799, "finished": true, "Ev Idx": 1185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564828253.016, "dur": 2.696,
    "args": {
      "External id": 1187,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 799, "finished": true, "Ev Idx": 1186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564828254.180, "dur": 1.062,
    "args": {
      "External id": 1188,"Record function id": 0, "finished": true, "Ev Idx": 1187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828254.693, "dur": 0.423,
    "args": {
      "External id": 1189,"Record function id": 0, "finished": true, "Ev Idx": 1188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564828256.029, "dur": 41.010,
    "args": {
      "External id": 1190,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 800, "finished": true, "Ev Idx": 1189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828256.612, "dur": 1.895,
    "args": {
      "External id": 1191,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 800, "finished": true, "Ev Idx": 1190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828256.994, "dur": 1.408,
    "args": {
      "External id": 1192,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 800, "finished": true, "Ev Idx": 1191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564828258.819, "dur": 36.213,
    "args": {
      "External id": 1193,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 801, "finished": true, "Ev Idx": 1192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564828271.078, "dur": 1.112,
    "args": {
      "External id": 1194,"Record function id": 0, "finished": true, "Ev Idx": 1193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564828295.651, "dur": 1.002,
    "args": {
      "External id": 1195,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 802, "finished": true, "Ev Idx": 1194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828299.611, "dur": 1.349,
    "args": {
      "External id": 1196,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 803, "finished": true, "Ev Idx": 1195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564828302.723, "dur": 2.028,
    "args": {
      "External id": 1197,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 804, "finished": true, "Ev Idx": 1196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828303.626, "dur": 0.568,
    "args": {
      "External id": 1198,"Record function id": 0, "finished": true, "Ev Idx": 1197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564828312.272, "dur": 40.250,
    "args": {
      "External id": 1199,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 805, "finished": true, "Ev Idx": 1198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564828312.905, "dur": 2.074,
    "args": {
      "External id": 1200,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 805, "finished": true, "Ev Idx": 1199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564828313.771, "dur": 0.800,
    "args": {
      "External id": 1201,"Record function id": 0, "finished": true, "Ev Idx": 1200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828314.171, "dur": 0.289,
    "args": {
      "External id": 1202,"Record function id": 0, "finished": true, "Ev Idx": 1201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564828315.289, "dur": 37.123,
    "args": {
      "External id": 1203,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 806, "finished": true, "Ev Idx": 1202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828315.682, "dur": 2.522,
    "args": {
      "External id": 1204,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 806, "finished": true, "Ev Idx": 1203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828315.957, "dur": 2.156,
    "args": {
      "External id": 1205,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 806, "finished": true, "Ev Idx": 1204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564828318.519, "dur": 31.974,
    "args": {
      "External id": 1206,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 807, "finished": true, "Ev Idx": 1205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564828330.635, "dur": 1.068,
    "args": {
      "External id": 1207,"Record function id": 0, "finished": true, "Ev Idx": 1206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564828351.086, "dur": 1.026,
    "args": {
      "External id": 1208,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 808, "finished": true, "Ev Idx": 1207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828354.740, "dur": 1.149,
    "args": {
      "External id": 1209,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 809, "finished": true, "Ev Idx": 1208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564828357.328, "dur": 1.995,
    "args": {
      "External id": 1210,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 810, "finished": true, "Ev Idx": 1209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828358.104, "dur": 0.574,
    "args": {
      "External id": 1211,"Record function id": 0, "finished": true, "Ev Idx": 1210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564828362.151, "dur": 2.151,
    "args": {
      "External id": 1212,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 811, "finished": true, "Ev Idx": 1211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828363.476, "dur": 0.319,
    "args": {
      "External id": 1213,"Record function id": 0, "finished": true, "Ev Idx": 1212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564828365.443, "dur": 1.044,
    "args": {
      "External id": 1214,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 811, "finished": true, "Ev Idx": 1213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828365.878, "dur": 0.268,
    "args": {
      "External id": 1215,"Record function id": 0, "finished": true, "Ev Idx": 1214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564828367.469, "dur": 12.294,
    "args": {
      "External id": 1216,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 811, "finished": true, "Ev Idx": 1215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828384.939, "dur": 4.719,
    "args": {
      "External id": 1217,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 812, "finished": true, "Ev Idx": 1216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828387.831, "dur": 0.571,
    "args": {
      "External id": 1218,"Record function id": 0, "finished": true, "Ev Idx": 1217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828391.757, "dur": 1.794,
    "args": {
      "External id": 1219,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 813, "finished": true, "Ev Idx": 1218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828392.645, "dur": 0.322,
    "args": {
      "External id": 1220,"Record function id": 0, "finished": true, "Ev Idx": 1219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564828395.474, "dur": 10.831,
    "args": {
      "External id": 1221,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 814, "finished": true, "Ev Idx": 1220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564828410.522, "dur": 17.623,
    "args": {
      "External id": 1222,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 815, "finished": true, "Ev Idx": 1221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564828431.061, "dur": 7.239,
    "args": {
      "External id": 1223,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 816, "finished": true, "Ev Idx": 1222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564828439.991, "dur": 9.648,
    "args": {
      "External id": 1224,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 817, "finished": true, "Ev Idx": 1223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564828451.744, "dur": 6.766,
    "args": {
      "External id": 1225,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 818, "finished": true, "Ev Idx": 1224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828461.203, "dur": 2.351,
    "args": {
      "External id": 1226,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 819, "finished": true, "Ev Idx": 1225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828462.379, "dur": 0.484,
    "args": {
      "External id": 1227,"Record function id": 0, "finished": true, "Ev Idx": 1226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828465.129, "dur": 1.787,
    "args": {
      "External id": 1228,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 820, "finished": true, "Ev Idx": 1227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828465.994, "dur": 0.350,
    "args": {
      "External id": 1229,"Record function id": 0, "finished": true, "Ev Idx": 1228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564828467.963, "dur": 9.115,
    "args": {
      "External id": 1230,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 821, "finished": true, "Ev Idx": 1229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564828478.700, "dur": 9.776,
    "args": {
      "External id": 1231,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 822, "finished": true, "Ev Idx": 1230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564828490.591, "dur": 7.667,
    "args": {
      "External id": 1232,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 823, "finished": true, "Ev Idx": 1231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564828499.922, "dur": 6.656,
    "args": {
      "External id": 1233,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 824, "finished": true, "Ev Idx": 1232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828529.788, "dur": 2.436,
    "args": {
      "External id": 1234,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 825, "finished": true, "Ev Idx": 1233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828531.022, "dur": 0.458,
    "args": {
      "External id": 1235,"Record function id": 0, "finished": true, "Ev Idx": 1234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828532.883, "dur": 1.710,
    "args": {
      "External id": 1236,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 826, "finished": true, "Ev Idx": 1235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828533.772, "dur": 0.323,
    "args": {
      "External id": 1237,"Record function id": 0, "finished": true, "Ev Idx": 1236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564828535.338, "dur": 3.139,
    "args": {
      "External id": 1238,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 827, "finished": true, "Ev Idx": 1237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828537.302, "dur": 0.596,
    "args": {
      "External id": 1239,"Record function id": 0, "finished": true, "Ev Idx": 1238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828539.233, "dur": 2.555,
    "args": {
      "External id": 1240,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 828, "finished": true, "Ev Idx": 1239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828540.987, "dur": 0.324,
    "args": {
      "External id": 1241,"Record function id": 0, "finished": true, "Ev Idx": 1240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828542.374, "dur": 1.747,
    "args": {
      "External id": 1242,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 829, "finished": true, "Ev Idx": 1241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828543.380, "dur": 0.271,
    "args": {
      "External id": 1243,"Record function id": 0, "finished": true, "Ev Idx": 1242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564828546.780, "dur": 3.031,
    "args": {
      "External id": 1244,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 830, "finished": true, "Ev Idx": 1243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828548.626, "dur": 0.525,
    "args": {
      "External id": 1245,"Record function id": 0, "finished": true, "Ev Idx": 1244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828552.184, "dur": 20.277,
    "args": {
      "External id": 1246,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 831, "finished": true, "Ev Idx": 1245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564828554.804, "dur": 15.359,
    "args": {
      "External id": 1247,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 831, "finished": true, "Ev Idx": 1246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564828556.291, "dur": 2.748,
    "args": {
      "External id": 1248,"Record function id": 0, "finished": true, "Ev Idx": 1247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564828557.224, "dur": 1.624,
    "args": {
      "External id": 1249,"Record function id": 0, "finished": true, "Ev Idx": 1248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564828559.761, "dur": 9.770,
    "args": {
      "External id": 1250,"Record function id": 0, "finished": true, "Ev Idx": 1249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564828570.641, "dur": 1.585,
    "args": {
      "External id": 1251,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 832, "finished": true, "Ev Idx": 1250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828575.600, "dur": 2.737,
    "args": {
      "External id": 1252,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 833, "finished": true, "Ev Idx": 1251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828577.438, "dur": 0.366,
    "args": {
      "External id": 1253,"Record function id": 0, "finished": true, "Ev Idx": 1252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828579.079, "dur": 2.024,
    "args": {
      "External id": 1254,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 834, "finished": true, "Ev Idx": 1253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828580.353, "dur": 0.275,
    "args": {
      "External id": 1255,"Record function id": 0, "finished": true, "Ev Idx": 1254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564828581.712, "dur": 1.475,
    "args": {
      "External id": 1256,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 835, "finished": true, "Ev Idx": 1255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828582.436, "dur": 0.298,
    "args": {
      "External id": 1257,"Record function id": 0, "finished": true, "Ev Idx": 1256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828583.746, "dur": 1.450,
    "args": {
      "External id": 1258,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 836, "finished": true, "Ev Idx": 1257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828584.418, "dur": 0.289,
    "args": {
      "External id": 1259,"Record function id": 0, "finished": true, "Ev Idx": 1258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828585.770, "dur": 1.597,
    "args": {
      "External id": 1260,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 837, "finished": true, "Ev Idx": 1259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828586.621, "dur": 0.271,
    "args": {
      "External id": 1261,"Record function id": 0, "finished": true, "Ev Idx": 1260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564828588.935, "dur": 2.052,
    "args": {
      "External id": 1262,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 838, "finished": true, "Ev Idx": 1261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828590.236, "dur": 0.305,
    "args": {
      "External id": 1263,"Record function id": 0, "finished": true, "Ev Idx": 1262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828592.348, "dur": 11.057,
    "args": {
      "External id": 1264,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 839, "finished": true, "Ev Idx": 1263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564828593.267, "dur": 8.720,
    "args": {
      "External id": 1265,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 839, "finished": true, "Ev Idx": 1264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564828594.516, "dur": 1.616,
    "args": {
      "External id": 1266,"Record function id": 0, "finished": true, "Ev Idx": 1265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564828594.832, "dur": 1.174,
    "args": {
      "External id": 1267,"Record function id": 0, "finished": true, "Ev Idx": 1266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564828596.427, "dur": 5.239,
    "args": {
      "External id": 1268,"Record function id": 0, "finished": true, "Ev Idx": 1267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564828602.356, "dur": 0.897,
    "args": {
      "External id": 1269,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 840, "finished": true, "Ev Idx": 1268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564828605.389, "dur": 1.869,
    "args": {
      "External id": 1270,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 841, "finished": true, "Ev Idx": 1269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828606.395, "dur": 0.381,
    "args": {
      "External id": 1271,"Record function id": 0, "finished": true, "Ev Idx": 1270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564828608.733, "dur": 45.339,
    "args": {
      "External id": 1272,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 842, "finished": true, "Ev Idx": 1271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564828609.347, "dur": 1.771,
    "args": {
      "External id": 1273,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 842, "finished": true, "Ev Idx": 1272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828610.338, "dur": 0.336,
    "args": {
      "External id": 1274,"Record function id": 0, "finished": true, "Ev Idx": 1273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828611.455, "dur": 3.783,
    "args": {
      "External id": 1275,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 843, "finished": true, "Ev Idx": 1274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564828613.283, "dur": 1.822,
    "args": {
      "External id": 1276,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 843, "finished": true, "Ev Idx": 1275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564828615.703, "dur": 2.353,
    "args": {
      "External id": 1277,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 844, "finished": true, "Ev Idx": 1276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828617.365, "dur": 0.312,
    "args": {
      "External id": 1278,"Record function id": 0, "finished": true, "Ev Idx": 1277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828618.339, "dur": 2.431,
    "args": {
      "External id": 1279,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 845, "finished": true, "Ev Idx": 1278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564828619.316, "dur": 1.340,
    "args": {
      "External id": 1280,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 845, "finished": true, "Ev Idx": 1279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564828621.498, "dur": 30.532,
    "args": {
      "External id": 1281,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 846, "finished": true, "Ev Idx": 1280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564828633.607, "dur": 1.041,
    "args": {
      "External id": 1282,"Record function id": 0, "finished": true, "Ev Idx": 1281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564828652.635, "dur": 1.089,
    "args": {
      "External id": 1283,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 847, "finished": true, "Ev Idx": 1282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564828657.350, "dur": 8.944,
    "args": {
      "External id": 1284,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 848, "finished": true, "Ev Idx": 1283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828669.337, "dur": 1.858,
    "args": {
      "External id": 1285,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 849, "finished": true, "Ev Idx": 1284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828670.164, "dur": 0.435,
    "args": {
      "External id": 1286,"Record function id": 0, "finished": true, "Ev Idx": 1285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828671.933, "dur": 1.238,
    "args": {
      "External id": 1287,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 849, "finished": true, "Ev Idx": 1286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828672.503, "dur": 0.273,
    "args": {
      "External id": 1288,"Record function id": 0, "finished": true, "Ev Idx": 1287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564828673.793, "dur": 1.143,
    "args": {
      "External id": 1289,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 849, "finished": true, "Ev Idx": 1288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828674.414, "dur": 0.195,
    "args": {
      "External id": 1290,"Record function id": 0, "finished": true, "Ev Idx": 1289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564828676.502, "dur": 7.098,
    "args": {
      "External id": 1291,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 849, "finished": true, "Ev Idx": 1290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564828688.889, "dur": 13.067,
    "args": {
      "External id": 1292,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 850, "finished": true, "Ev Idx": 1291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564828689.795, "dur": 0.408,
    "args": {
      "External id": 1293,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 850, "finished": true, "Ev Idx": 1292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564828690.828, "dur": 10.858,
    "args": {
      "External id": 1294,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 850, "finished": true, "Ev Idx": 1293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564828704.806, "dur": 0.179,
    "args": {
      "External id": 1295,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 851, "finished": true, "Ev Idx": 1294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564828716.265, "dur": 0.411,
    "args": {
      "External id": 1296,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 851, "finished": true, "Ev Idx": 1295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564828718.358, "dur": 40.107,
    "args": {
      "External id": 1297,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 851, "finished": true, "Ev Idx": 1296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564828718.856, "dur": 2.597,
    "args": {
      "External id": 1298,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 851, "finished": true, "Ev Idx": 1297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828720.584, "dur": 0.374,
    "args": {
      "External id": 1299,"Record function id": 0, "finished": true, "Ev Idx": 1298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828721.827, "dur": 2.861,
    "args": {
      "External id": 1300,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 852, "finished": true, "Ev Idx": 1299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828722.293, "dur": 2.272,
    "args": {
      "External id": 1301,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 852, "finished": true, "Ev Idx": 1300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564828725.260, "dur": 1.247,
    "args": {
      "External id": 1302,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 853, "finished": true, "Ev Idx": 1301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828725.874, "dur": 0.274,
    "args": {
      "External id": 1303,"Record function id": 0, "finished": true, "Ev Idx": 1302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828726.786, "dur": 1.707,
    "args": {
      "External id": 1304,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 854, "finished": true, "Ev Idx": 1303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828727.048, "dur": 1.386,
    "args": {
      "External id": 1305,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 854, "finished": true, "Ev Idx": 1304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564828728.866, "dur": 27.480,
    "args": {
      "External id": 1306,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 855, "finished": true, "Ev Idx": 1305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564828740.123, "dur": 1.180,
    "args": {
      "External id": 1307,"Record function id": 0, "finished": true, "Ev Idx": 1306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564828757.016, "dur": 1.159,
    "args": {
      "External id": 1308,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 856, "finished": true, "Ev Idx": 1307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564828760.496, "dur": 1.699,
    "args": {
      "External id": 1309,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 857, "finished": true, "Ev Idx": 1308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828761.324, "dur": 0.377,
    "args": {
      "External id": 1310,"Record function id": 0, "finished": true, "Ev Idx": 1309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564828764.626, "dur": 11.623,
    "args": {
      "External id": 1311,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 858, "finished": true, "Ev Idx": 1310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564828765.215, "dur": 10.894,
    "args": {
      "External id": 1312,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 858, "finished": true, "Ev Idx": 1311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564828766.230, "dur": 1.863,
    "args": {
      "External id": 1313,"Record function id": 0, "finished": true, "Ev Idx": 1312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564828766.599, "dur": 1.348,
    "args": {
      "External id": 1314,"Record function id": 0, "finished": true, "Ev Idx": 1313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564828768.508, "dur": 7.191,
    "args": {
      "External id": 1315,"Record function id": 0, "finished": true, "Ev Idx": 1314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828781.773, "dur": 1.774,
    "args": {
      "External id": 1316,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 859, "finished": true, "Ev Idx": 1315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828782.243, "dur": 1.207,
    "args": {
      "External id": 1317,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 859, "finished": true, "Ev Idx": 1316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564828795.899, "dur": 35.867,
    "args": {
      "External id": 1318,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 860, "finished": true, "Ev Idx": 1317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564828796.423, "dur": 3.107,
    "args": {
      "External id": 1319,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 860, "finished": true, "Ev Idx": 1318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564828797.885, "dur": 1.132,
    "args": {
      "External id": 1320,"Record function id": 0, "finished": true, "Ev Idx": 1319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828798.569, "dur": 0.331,
    "args": {
      "External id": 1321,"Record function id": 0, "finished": true, "Ev Idx": 1320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564828800.001, "dur": 31.626,
    "args": {
      "External id": 1322,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 861, "finished": true, "Ev Idx": 1321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828800.413, "dur": 2.323,
    "args": {
      "External id": 1323,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 861, "finished": true, "Ev Idx": 1322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828800.675, "dur": 1.973,
    "args": {
      "External id": 1324,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 861, "finished": true, "Ev Idx": 1323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564828803.080, "dur": 26.648,
    "args": {
      "External id": 1325,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 862, "finished": true, "Ev Idx": 1324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564828814.083, "dur": 1.161,
    "args": {
      "External id": 1326,"Record function id": 0, "finished": true, "Ev Idx": 1325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564828830.315, "dur": 0.983,
    "args": {
      "External id": 1327,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 863, "finished": true, "Ev Idx": 1326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564828836.399, "dur": 8.301,
    "args": {
      "External id": 1328,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 864, "finished": true, "Ev Idx": 1327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564828854.064, "dur": 0.170,
    "args": {
      "External id": 1329,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 865, "finished": true, "Ev Idx": 1328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564828856.279, "dur": 10.440,
    "args": {
      "External id": 1330,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 865, "finished": true, "Ev Idx": 1329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564828857.930, "dur": 0.398,
    "args": {
      "External id": 1331,"Record function id": 0, "finished": true, "Ev Idx": 1330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564828858.730, "dur": 0.295,
    "args": {
      "External id": 1332,"Record function id": 0, "finished": true, "Ev Idx": 1331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564828869.406, "dur": 14.175,
    "args": {
      "External id": 1333,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 866, "finished": true, "Ev Idx": 1332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564828887.025, "dur": 8.544,
    "args": {
      "External id": 1334,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 867, "finished": true, "Ev Idx": 1333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564828897.412, "dur": 8.600,
    "args": {
      "External id": 1335,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 868, "finished": true, "Ev Idx": 1334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564828907.851, "dur": 11.177,
    "args": {
      "External id": 1336,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 869, "finished": true, "Ev Idx": 1335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564828921.417, "dur": 0.262,
    "args": {
      "External id": 1337,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 870, "finished": true, "Ev Idx": 1336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564828922.624, "dur": 8.596,
    "args": {
      "External id": 1338,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 870, "finished": true, "Ev Idx": 1337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564828944.815, "dur": 110.461,
    "args": {
      "External id": 1339,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 871, "finished": true, "Ev Idx": 1338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564828945.378, "dur": 2.978,
    "args": {
      "External id": 1340,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 871, "finished": true, "Ev Idx": 1339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564828946.479, "dur": 1.296,
    "args": {
      "External id": 1341,"Record function id": 0, "finished": true, "Ev Idx": 1340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564828947.142, "dur": 0.520,
    "args": {
      "External id": 1342,"Record function id": 0, "finished": true, "Ev Idx": 1341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564828948.843, "dur": 106.273,
    "args": {
      "External id": 1343,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 872, "finished": true, "Ev Idx": 1342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564828949.269, "dur": 3.433,
    "args": {
      "External id": 1344,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 872, "finished": true, "Ev Idx": 1343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564828949.638, "dur": 2.939,
    "args": {
      "External id": 1345,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 872, "finished": true, "Ev Idx": 1344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564828952.991, "dur": 99.385,
    "args": {
      "External id": 1346,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 873, "finished": true, "Ev Idx": 1345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829022.497, "dur": 1.281,
    "args": {
      "External id": 1347,"Record function id": 0, "finished": true, "Ev Idx": 1346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564829053.186, "dur": 1.483,
    "args": {
      "External id": 1348,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 874, "finished": true, "Ev Idx": 1347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564829065.977, "dur": 11.134,
    "args": {
      "External id": 1349,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 875, "finished": true, "Ev Idx": 1348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564829085.792, "dur": 80.588,
    "args": {
      "External id": 1350,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 876, "finished": true, "Ev Idx": 1349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564829086.129, "dur": 3.495,
    "args": {
      "External id": 1351,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 876, "finished": true, "Ev Idx": 1350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564829087.475, "dur": 1.541,
    "args": {
      "External id": 1352,"Record function id": 0, "finished": true, "Ev Idx": 1351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829088.220, "dur": 0.637,
    "args": {
      "External id": 1353,"Record function id": 0, "finished": true, "Ev Idx": 1352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564829090.062, "dur": 76.217,
    "args": {
      "External id": 1354,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 877, "finished": true, "Ev Idx": 1353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829090.527, "dur": 3.440,
    "args": {
      "External id": 1355,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 877, "finished": true, "Ev Idx": 1354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564829090.868, "dur": 3.001,
    "args": {
      "External id": 1356,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 877, "finished": true, "Ev Idx": 1355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564829094.300, "dur": 69.780,
    "args": {
      "External id": 1357,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 878, "finished": true, "Ev Idx": 1356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829139.023, "dur": 1.100,
    "args": {
      "External id": 1358,"Record function id": 0, "finished": true, "Ev Idx": 1357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564829164.699, "dur": 1.199,
    "args": {
      "External id": 1359,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 879, "finished": true, "Ev Idx": 1358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564829168.747, "dur": 9.492,
    "args": {
      "External id": 1360,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 880, "finished": true, "Ev Idx": 1359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564829186.042, "dur": 44.352,
    "args": {
      "External id": 1361,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 881, "finished": true, "Ev Idx": 1360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564829186.715, "dur": 2.537,
    "args": {
      "External id": 1362,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 881, "finished": true, "Ev Idx": 1361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564829187.551, "dur": 1.088,
    "args": {
      "External id": 1363,"Record function id": 0, "finished": true, "Ev Idx": 1362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829188.003, "dur": 0.500,
    "args": {
      "External id": 1364,"Record function id": 0, "finished": true, "Ev Idx": 1363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564829189.581, "dur": 40.717,
    "args": {
      "External id": 1365,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 882, "finished": true, "Ev Idx": 1364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829190.043, "dur": 2.070,
    "args": {
      "External id": 1366,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 882, "finished": true, "Ev Idx": 1365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564829190.501, "dur": 1.496,
    "args": {
      "External id": 1367,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 882, "finished": true, "Ev Idx": 1366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564829192.435, "dur": 35.987,
    "args": {
      "External id": 1368,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 883, "finished": true, "Ev Idx": 1367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829206.276, "dur": 1.175,
    "args": {
      "External id": 1369,"Record function id": 0, "finished": true, "Ev Idx": 1368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564829229.003, "dur": 0.958,
    "args": {
      "External id": 1370,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 884, "finished": true, "Ev Idx": 1369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564829233.088, "dur": 8.625,
    "args": {
      "External id": 1371,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 885, "finished": true, "Ev Idx": 1370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564829262.386, "dur": 0.189,
    "args": {
      "External id": 1372,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 886, "finished": true, "Ev Idx": 1371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564829264.656, "dur": 10.046,
    "args": {
      "External id": 1373,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 886, "finished": true, "Ev Idx": 1372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564829266.071, "dur": 0.489,
    "args": {
      "External id": 1374,"Record function id": 0, "finished": true, "Ev Idx": 1373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564829266.863, "dur": 0.253,
    "args": {
      "External id": 1375,"Record function id": 0, "finished": true, "Ev Idx": 1374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564829277.532, "dur": 14.084,
    "args": {
      "External id": 1376,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 887, "finished": true, "Ev Idx": 1375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564829295.309, "dur": 9.034,
    "args": {
      "External id": 1377,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 888, "finished": true, "Ev Idx": 1376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564829306.217, "dur": 8.648,
    "args": {
      "External id": 1378,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 889, "finished": true, "Ev Idx": 1377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564829316.527, "dur": 8.975,
    "args": {
      "External id": 1379,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 890, "finished": true, "Ev Idx": 1378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564829327.989, "dur": 0.095,
    "args": {
      "External id": 1380,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 891, "finished": true, "Ev Idx": 1379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564829329.068, "dur": 7.740,
    "args": {
      "External id": 1381,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 891, "finished": true, "Ev Idx": 1380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564829356.221, "dur": 38.911,
    "args": {
      "External id": 1382,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 892, "finished": true, "Ev Idx": 1381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564829356.656, "dur": 3.809,
    "args": {
      "External id": 1383,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 892, "finished": true, "Ev Idx": 1382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564829358.663, "dur": 1.202,
    "args": {
      "External id": 1384,"Record function id": 0, "finished": true, "Ev Idx": 1383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829359.260, "dur": 0.462,
    "args": {
      "External id": 1385,"Record function id": 0, "finished": true, "Ev Idx": 1384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564829360.818, "dur": 34.186,
    "args": {
      "External id": 1386,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 893, "finished": true, "Ev Idx": 1385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829361.427, "dur": 1.685,
    "args": {
      "External id": 1387,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 893, "finished": true, "Ev Idx": 1386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564829361.733, "dur": 1.280,
    "args": {
      "External id": 1388,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 893, "finished": true, "Ev Idx": 1387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564829363.452, "dur": 29.448,
    "args": {
      "External id": 1389,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 894, "finished": true, "Ev Idx": 1388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829375.549, "dur": 1.191,
    "args": {
      "External id": 1390,"Record function id": 0, "finished": true, "Ev Idx": 1389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564829393.552, "dur": 1.134,
    "args": {
      "External id": 1391,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 895, "finished": true, "Ev Idx": 1390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564829398.271, "dur": 1.473,
    "args": {
      "External id": 1392,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 896, "finished": true, "Ev Idx": 1391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564829401.639, "dur": 2.683,
    "args": {
      "External id": 1393,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 897, "finished": true, "Ev Idx": 1392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829402.702, "dur": 0.721,
    "args": {
      "External id": 1394,"Record function id": 0, "finished": true, "Ev Idx": 1393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564829411.850, "dur": 40.801,
    "args": {
      "External id": 1395,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 898, "finished": true, "Ev Idx": 1394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564829412.345, "dur": 2.250,
    "args": {
      "External id": 1396,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 898, "finished": true, "Ev Idx": 1395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564829413.239, "dur": 0.896,
    "args": {
      "External id": 1397,"Record function id": 0, "finished": true, "Ev Idx": 1396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829413.748, "dur": 0.296,
    "args": {
      "External id": 1398,"Record function id": 0, "finished": true, "Ev Idx": 1397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564829414.972, "dur": 37.582,
    "args": {
      "External id": 1399,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 899, "finished": true, "Ev Idx": 1398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829415.355, "dur": 1.244,
    "args": {
      "External id": 1400,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 899, "finished": true, "Ev Idx": 1399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564829415.614, "dur": 0.873,
    "args": {
      "External id": 1401,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 899, "finished": true, "Ev Idx": 1400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564829416.932, "dur": 33.634,
    "args": {
      "External id": 1402,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 900, "finished": true, "Ev Idx": 1401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829427.683, "dur": 1.135,
    "args": {
      "External id": 1403,"Record function id": 0, "finished": true, "Ev Idx": 1402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564829451.123, "dur": 1.112,
    "args": {
      "External id": 1404,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 901, "finished": true, "Ev Idx": 1403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564829455.083, "dur": 1.228,
    "args": {
      "External id": 1405,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 902, "finished": true, "Ev Idx": 1404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564829457.687, "dur": 2.032,
    "args": {
      "External id": 1406,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 903, "finished": true, "Ev Idx": 1405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829458.745, "dur": 0.451,
    "args": {
      "External id": 1407,"Record function id": 0, "finished": true, "Ev Idx": 1406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564829466.989, "dur": 38.040,
    "args": {
      "External id": 1408,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 904, "finished": true, "Ev Idx": 1407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564829467.292, "dur": 2.210,
    "args": {
      "External id": 1409,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 904, "finished": true, "Ev Idx": 1408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564829468.125, "dur": 0.963,
    "args": {
      "External id": 1410,"Record function id": 0, "finished": true, "Ev Idx": 1409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829468.721, "dur": 0.271,
    "args": {
      "External id": 1411,"Record function id": 0, "finished": true, "Ev Idx": 1410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564829469.957, "dur": 34.974,
    "args": {
      "External id": 1412,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 905, "finished": true, "Ev Idx": 1411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829470.312, "dur": 1.683,
    "args": {
      "External id": 1413,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 905, "finished": true, "Ev Idx": 1412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564829470.894, "dur": 1.008,
    "args": {
      "External id": 1414,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 905, "finished": true, "Ev Idx": 1413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564829472.313, "dur": 30.731,
    "args": {
      "External id": 1415,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 906, "finished": true, "Ev Idx": 1414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829483.319, "dur": 1.037,
    "args": {
      "External id": 1416,"Record function id": 0, "finished": true, "Ev Idx": 1415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564829503.575, "dur": 1.101,
    "args": {
      "External id": 1417,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 907, "finished": true, "Ev Idx": 1416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564829507.725, "dur": 1.134,
    "args": {
      "External id": 1418,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 908, "finished": true, "Ev Idx": 1417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564829510.341, "dur": 1.719,
    "args": {
      "External id": 1419,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 909, "finished": true, "Ev Idx": 1418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829511.016, "dur": 0.560,
    "args": {
      "External id": 1420,"Record function id": 0, "finished": true, "Ev Idx": 1419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564829514.844, "dur": 2.222,
    "args": {
      "External id": 1421,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 910, "finished": true, "Ev Idx": 1420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829516.050, "dur": 0.470,
    "args": {
      "External id": 1422,"Record function id": 0, "finished": true, "Ev Idx": 1421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564829518.206, "dur": 1.527,
    "args": {
      "External id": 1423,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 910, "finished": true, "Ev Idx": 1422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829518.803, "dur": 0.425,
    "args": {
      "External id": 1424,"Record function id": 0, "finished": true, "Ev Idx": 1423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564829520.775, "dur": 9.988,
    "args": {
      "External id": 1425,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 910, "finished": true, "Ev Idx": 1424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829536.331, "dur": 5.143,
    "args": {
      "External id": 1426,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 911, "finished": true, "Ev Idx": 1425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829538.784, "dur": 0.710,
    "args": {
      "External id": 1427,"Record function id": 0, "finished": true, "Ev Idx": 1426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829543.761, "dur": 2.361,
    "args": {
      "External id": 1428,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 912, "finished": true, "Ev Idx": 1427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829545.203, "dur": 0.346,
    "args": {
      "External id": 1429,"Record function id": 0, "finished": true, "Ev Idx": 1428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564829548.032, "dur": 9.996,
    "args": {
      "External id": 1430,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 913, "finished": true, "Ev Idx": 1429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564829562.125, "dur": 18.298,
    "args": {
      "External id": 1431,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 914, "finished": true, "Ev Idx": 1430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564829583.283, "dur": 51.512,
    "args": {
      "External id": 1432,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 915, "finished": true, "Ev Idx": 1431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564829637.061, "dur": 8.688,
    "args": {
      "External id": 1433,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 916, "finished": true, "Ev Idx": 1432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564829647.969, "dur": 6.709,
    "args": {
      "External id": 1434,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 917, "finished": true, "Ev Idx": 1433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829657.970, "dur": 2.851,
    "args": {
      "External id": 1435,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 918, "finished": true, "Ev Idx": 1434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829659.414, "dur": 0.493,
    "args": {
      "External id": 1436,"Record function id": 0, "finished": true, "Ev Idx": 1435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829662.605, "dur": 2.140,
    "args": {
      "External id": 1437,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 919, "finished": true, "Ev Idx": 1436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829663.692, "dur": 0.501,
    "args": {
      "External id": 1438,"Record function id": 0, "finished": true, "Ev Idx": 1437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564829665.911, "dur": 7.594,
    "args": {
      "External id": 1439,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 920, "finished": true, "Ev Idx": 1438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564829675.922, "dur": 10.649,
    "args": {
      "External id": 1440,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 921, "finished": true, "Ev Idx": 1439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564829689.022, "dur": 7.368,
    "args": {
      "External id": 1441,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 922, "finished": true, "Ev Idx": 1440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564829697.712, "dur": 6.630,
    "args": {
      "External id": 1442,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 923, "finished": true, "Ev Idx": 1441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829728.931, "dur": 2.601,
    "args": {
      "External id": 1443,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 924, "finished": true, "Ev Idx": 1442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829730.253, "dur": 0.532,
    "args": {
      "External id": 1444,"Record function id": 0, "finished": true, "Ev Idx": 1443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829732.219, "dur": 1.563,
    "args": {
      "External id": 1445,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 925, "finished": true, "Ev Idx": 1444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829732.959, "dur": 0.306,
    "args": {
      "External id": 1446,"Record function id": 0, "finished": true, "Ev Idx": 1445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564829734.634, "dur": 3.543,
    "args": {
      "External id": 1447,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 926, "finished": true, "Ev Idx": 1446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829736.632, "dur": 0.892,
    "args": {
      "External id": 1448,"Record function id": 0, "finished": true, "Ev Idx": 1447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829738.927, "dur": 2.606,
    "args": {
      "External id": 1449,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 927, "finished": true, "Ev Idx": 1448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829740.668, "dur": 0.359,
    "args": {
      "External id": 1450,"Record function id": 0, "finished": true, "Ev Idx": 1449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829742.251, "dur": 1.705,
    "args": {
      "External id": 1451,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 928, "finished": true, "Ev Idx": 1450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829743.223, "dur": 0.271,
    "args": {
      "External id": 1452,"Record function id": 0, "finished": true, "Ev Idx": 1451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564829747.008, "dur": 3.200,
    "args": {
      "External id": 1453,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 929, "finished": true, "Ev Idx": 1452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829749.038, "dur": 0.405,
    "args": {
      "External id": 1454,"Record function id": 0, "finished": true, "Ev Idx": 1453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829752.782, "dur": 21.117,
    "args": {
      "External id": 1455,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 930, "finished": true, "Ev Idx": 1454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564829755.474, "dur": 15.768,
    "args": {
      "External id": 1456,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 930, "finished": true, "Ev Idx": 1455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564829757.293, "dur": 2.847,
    "args": {
      "External id": 1457,"Record function id": 0, "finished": true, "Ev Idx": 1456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829758.106, "dur": 1.805,
    "args": {
      "External id": 1458,"Record function id": 0, "finished": true, "Ev Idx": 1457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564829760.844, "dur": 9.802,
    "args": {
      "External id": 1459,"Record function id": 0, "finished": true, "Ev Idx": 1458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564829771.845, "dur": 1.789,
    "args": {
      "External id": 1460,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 931, "finished": true, "Ev Idx": 1459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829777.206, "dur": 2.806,
    "args": {
      "External id": 1461,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 932, "finished": true, "Ev Idx": 1460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829779.019, "dur": 0.451,
    "args": {
      "External id": 1462,"Record function id": 0, "finished": true, "Ev Idx": 1461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829780.618, "dur": 1.757,
    "args": {
      "External id": 1463,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 933, "finished": true, "Ev Idx": 1462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829781.691, "dur": 0.273,
    "args": {
      "External id": 1464,"Record function id": 0, "finished": true, "Ev Idx": 1463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564829782.878, "dur": 1.448,
    "args": {
      "External id": 1465,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 934, "finished": true, "Ev Idx": 1464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829783.585, "dur": 0.310,
    "args": {
      "External id": 1466,"Record function id": 0, "finished": true, "Ev Idx": 1465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829784.929, "dur": 1.441,
    "args": {
      "External id": 1467,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 935, "finished": true, "Ev Idx": 1466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829785.615, "dur": 0.291,
    "args": {
      "External id": 1468,"Record function id": 0, "finished": true, "Ev Idx": 1467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829786.953, "dur": 1.525,
    "args": {
      "External id": 1469,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 936, "finished": true, "Ev Idx": 1468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829787.601, "dur": 0.444,
    "args": {
      "External id": 1470,"Record function id": 0, "finished": true, "Ev Idx": 1469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564829790.034, "dur": 2.217,
    "args": {
      "External id": 1471,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 937, "finished": true, "Ev Idx": 1470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829791.394, "dur": 0.464,
    "args": {
      "External id": 1472,"Record function id": 0, "finished": true, "Ev Idx": 1471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829793.651, "dur": 12.603,
    "args": {
      "External id": 1473,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 938, "finished": true, "Ev Idx": 1472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564829794.738, "dur": 9.872,
    "args": {
      "External id": 1474,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 938, "finished": true, "Ev Idx": 1473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564829796.437, "dur": 1.987,
    "args": {
      "External id": 1475,"Record function id": 0, "finished": true, "Ev Idx": 1474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829796.966, "dur": 1.352,
    "args": {
      "External id": 1476,"Record function id": 0, "finished": true, "Ev Idx": 1475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564829798.746, "dur": 5.515,
    "args": {
      "External id": 1477,"Record function id": 0, "finished": true, "Ev Idx": 1476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564829804.991, "dur": 1.087,
    "args": {
      "External id": 1478,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 939, "finished": true, "Ev Idx": 1477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564829808.738, "dur": 1.911,
    "args": {
      "External id": 1479,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 940, "finished": true, "Ev Idx": 1478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829809.695, "dur": 0.373,
    "args": {
      "External id": 1480,"Record function id": 0, "finished": true, "Ev Idx": 1479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564829812.349, "dur": 50.556,
    "args": {
      "External id": 1481,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 941, "finished": true, "Ev Idx": 1480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564829813.235, "dur": 1.644,
    "args": {
      "External id": 1482,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 941, "finished": true, "Ev Idx": 1481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829814.106, "dur": 0.327,
    "args": {
      "External id": 1483,"Record function id": 0, "finished": true, "Ev Idx": 1482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829815.211, "dur": 3.738,
    "args": {
      "External id": 1484,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 942, "finished": true, "Ev Idx": 1483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564829816.849, "dur": 1.968,
    "args": {
      "External id": 1485,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 942, "finished": true, "Ev Idx": 1484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564829819.422, "dur": 2.463,
    "args": {
      "External id": 1486,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 943, "finished": true, "Ev Idx": 1485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829821.066, "dur": 0.436,
    "args": {
      "External id": 1487,"Record function id": 0, "finished": true, "Ev Idx": 1486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829822.170, "dur": 2.424,
    "args": {
      "External id": 1488,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 944, "finished": true, "Ev Idx": 1487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564829823.015, "dur": 1.483,
    "args": {
      "External id": 1489,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 944, "finished": true, "Ev Idx": 1488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564829825.252, "dur": 35.500,
    "args": {
      "External id": 1490,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 945, "finished": true, "Ev Idx": 1489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829838.757, "dur": 1.167,
    "args": {
      "External id": 1491,"Record function id": 0, "finished": true, "Ev Idx": 1490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564829861.323, "dur": 1.140,
    "args": {
      "External id": 1492,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 946, "finished": true, "Ev Idx": 1491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564829866.292, "dur": 9.647,
    "args": {
      "External id": 1493,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 947, "finished": true, "Ev Idx": 1492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829879.062, "dur": 2.158,
    "args": {
      "External id": 1494,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 948, "finished": true, "Ev Idx": 1493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829880.212, "dur": 0.453,
    "args": {
      "External id": 1495,"Record function id": 0, "finished": true, "Ev Idx": 1494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829882.046, "dur": 1.480,
    "args": {
      "External id": 1496,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 948, "finished": true, "Ev Idx": 1495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829882.817, "dur": 0.309,
    "args": {
      "External id": 1497,"Record function id": 0, "finished": true, "Ev Idx": 1496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564829884.288, "dur": 1.353,
    "args": {
      "External id": 1498,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 948, "finished": true, "Ev Idx": 1497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829884.917, "dur": 0.399,
    "args": {
      "External id": 1499,"Record function id": 0, "finished": true, "Ev Idx": 1498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564829887.363, "dur": 8.796,
    "args": {
      "External id": 1500,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 948, "finished": true, "Ev Idx": 1499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564829901.875, "dur": 14.822,
    "args": {
      "External id": 1501,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 949, "finished": true, "Ev Idx": 1500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564829902.906, "dur": 0.572,
    "args": {
      "External id": 1502,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 949, "finished": true, "Ev Idx": 1501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564829904.142, "dur": 12.285,
    "args": {
      "External id": 1503,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 949, "finished": true, "Ev Idx": 1502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564829919.622, "dur": 0.529,
    "args": {
      "External id": 1504,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 950, "finished": true, "Ev Idx": 1503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564829931.900, "dur": 0.441,
    "args": {
      "External id": 1505,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 950, "finished": true, "Ev Idx": 1504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564829934.189, "dur": 41.676,
    "args": {
      "External id": 1506,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 950, "finished": true, "Ev Idx": 1505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564829934.685, "dur": 2.736,
    "args": {
      "External id": 1507,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 950, "finished": true, "Ev Idx": 1506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829936.459, "dur": 0.375,
    "args": {
      "External id": 1508,"Record function id": 0, "finished": true, "Ev Idx": 1507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829937.933, "dur": 2.845,
    "args": {
      "External id": 1509,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 951, "finished": true, "Ev Idx": 1508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564829938.278, "dur": 2.394,
    "args": {
      "External id": 1510,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 951, "finished": true, "Ev Idx": 1509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564829941.438, "dur": 1.215,
    "args": {
      "External id": 1511,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 952, "finished": true, "Ev Idx": 1510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829942.037, "dur": 0.266,
    "args": {
      "External id": 1512,"Record function id": 0, "finished": true, "Ev Idx": 1511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564829942.922, "dur": 1.379,
    "args": {
      "External id": 1513,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 953, "finished": true, "Ev Idx": 1512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564829943.333, "dur": 0.897,
    "args": {
      "External id": 1514,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 953, "finished": true, "Ev Idx": 1513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564829944.756, "dur": 29.184,
    "args": {
      "External id": 1515,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 954, "finished": true, "Ev Idx": 1514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829957.093, "dur": 1.247,
    "args": {
      "External id": 1516,"Record function id": 0, "finished": true, "Ev Idx": 1515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564829974.495, "dur": 1.017,
    "args": {
      "External id": 1517,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 955, "finished": true, "Ev Idx": 1516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564829978.015, "dur": 2.019,
    "args": {
      "External id": 1518,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 956, "finished": true, "Ev Idx": 1517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564829979.153, "dur": 0.354,
    "args": {
      "External id": 1519,"Record function id": 0, "finished": true, "Ev Idx": 1518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564829982.614, "dur": 11.772,
    "args": {
      "External id": 1520,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 957, "finished": true, "Ev Idx": 1519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564829983.185, "dur": 11.058,
    "args": {
      "External id": 1521,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 957, "finished": true, "Ev Idx": 1520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564829984.204, "dur": 2.380,
    "args": {
      "External id": 1522,"Record function id": 0, "finished": true, "Ev Idx": 1521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564829984.600, "dur": 1.852,
    "args": {
      "External id": 1523,"Record function id": 0, "finished": true, "Ev Idx": 1522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564829986.910, "dur": 6.878,
    "args": {
      "External id": 1524,"Record function id": 0, "finished": true, "Ev Idx": 1523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564830000.314, "dur": 1.650,
    "args": {
      "External id": 1525,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 958, "finished": true, "Ev Idx": 1524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830000.667, "dur": 1.194,
    "args": {
      "External id": 1526,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 958, "finished": true, "Ev Idx": 1525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564830016.129, "dur": 34.318,
    "args": {
      "External id": 1527,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 959, "finished": true, "Ev Idx": 1526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564830016.741, "dur": 3.234,
    "args": {
      "External id": 1528,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 959, "finished": true, "Ev Idx": 1527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564830018.370, "dur": 1.051,
    "args": {
      "External id": 1529,"Record function id": 0, "finished": true, "Ev Idx": 1528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830018.778, "dur": 0.515,
    "args": {
      "External id": 1530,"Record function id": 0, "finished": true, "Ev Idx": 1529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564830020.344, "dur": 29.962,
    "args": {
      "External id": 1531,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 960, "finished": true, "Ev Idx": 1530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564830020.974, "dur": 2.258,
    "args": {
      "External id": 1532,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 960, "finished": true, "Ev Idx": 1531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830021.403, "dur": 1.743,
    "args": {
      "External id": 1533,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 960, "finished": true, "Ev Idx": 1532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564830023.685, "dur": 24.589,
    "args": {
      "External id": 1534,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 961, "finished": true, "Ev Idx": 1533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564830033.287, "dur": 1.079,
    "args": {
      "External id": 1535,"Record function id": 0, "finished": true, "Ev Idx": 1534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564830048.933, "dur": 1.008,
    "args": {
      "External id": 1536,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 962, "finished": true, "Ev Idx": 1535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564830107.249, "dur": 13.137,
    "args": {
      "External id": 1537,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 963, "finished": true, "Ev Idx": 1536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564830132.638, "dur": 0.234,
    "args": {
      "External id": 1538,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 964, "finished": true, "Ev Idx": 1537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564830135.286, "dur": 12.988,
    "args": {
      "External id": 1539,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 964, "finished": true, "Ev Idx": 1538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564830137.731, "dur": 0.587,
    "args": {
      "External id": 1540,"Record function id": 0, "finished": true, "Ev Idx": 1539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564830138.823, "dur": 0.065,
    "args": {
      "External id": 1541,"Record function id": 0, "finished": true, "Ev Idx": 1540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564830151.201, "dur": 15.401,
    "args": {
      "External id": 1542,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 965, "finished": true, "Ev Idx": 1541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564830170.709, "dur": 8.744,
    "args": {
      "External id": 1543,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 966, "finished": true, "Ev Idx": 1542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564830181.484, "dur": 10.395,
    "args": {
      "External id": 1544,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 967, "finished": true, "Ev Idx": 1543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564830193.594, "dur": 8.590,
    "args": {
      "External id": 1545,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 968, "finished": true, "Ev Idx": 1544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564830204.802, "dur": 0.090,
    "args": {
      "External id": 1546,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 969, "finished": true, "Ev Idx": 1545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564830205.856, "dur": 9.582,
    "args": {
      "External id": 1547,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 969, "finished": true, "Ev Idx": 1546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564830231.033, "dur": 53.715,
    "args": {
      "External id": 1548,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 970, "finished": true, "Ev Idx": 1547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564830231.539, "dur": 3.232,
    "args": {
      "External id": 1549,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 970, "finished": true, "Ev Idx": 1548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564830232.695, "dur": 1.367,
    "args": {
      "External id": 1550,"Record function id": 0, "finished": true, "Ev Idx": 1549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830233.340, "dur": 0.582,
    "args": {
      "External id": 1551,"Record function id": 0, "finished": true, "Ev Idx": 1550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564830235.096, "dur": 49.509,
    "args": {
      "External id": 1552,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 971, "finished": true, "Ev Idx": 1551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564830235.769, "dur": 3.104,
    "args": {
      "External id": 1553,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 971, "finished": true, "Ev Idx": 1552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830236.154, "dur": 2.607,
    "args": {
      "External id": 1554,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 971, "finished": true, "Ev Idx": 1553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564830239.258, "dur": 43.233,
    "args": {
      "External id": 1555,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 972, "finished": true, "Ev Idx": 1554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564830251.806, "dur": 1.259,
    "args": {
      "External id": 1556,"Record function id": 0, "finished": true, "Ev Idx": 1555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564830283.145, "dur": 1.092,
    "args": {
      "External id": 1557,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 973, "finished": true, "Ev Idx": 1556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564830293.456, "dur": 70.893,
    "args": {
      "External id": 1558,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 974, "finished": true, "Ev Idx": 1557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564830373.318, "dur": 47.922,
    "args": {
      "External id": 1559,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 975, "finished": true, "Ev Idx": 1558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564830373.720, "dur": 4.728,
    "args": {
      "External id": 1560,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 975, "finished": true, "Ev Idx": 1559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564830376.240, "dur": 1.443,
    "args": {
      "External id": 1561,"Record function id": 0, "finished": true, "Ev Idx": 1560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830376.960, "dur": 0.565,
    "args": {
      "External id": 1562,"Record function id": 0, "finished": true, "Ev Idx": 1561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564830378.995, "dur": 42.147,
    "args": {
      "External id": 1563,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 976, "finished": true, "Ev Idx": 1562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564830379.397, "dur": 1.618,
    "args": {
      "External id": 1564,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 976, "finished": true, "Ev Idx": 1563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830379.712, "dur": 1.182,
    "args": {
      "External id": 1565,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 976, "finished": true, "Ev Idx": 1564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564830381.356, "dur": 37.904,
    "args": {
      "External id": 1566,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 977, "finished": true, "Ev Idx": 1565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564830395.107, "dur": 1.214,
    "args": {
      "External id": 1567,"Record function id": 0, "finished": true, "Ev Idx": 1566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564830419.912, "dur": 0.938,
    "args": {
      "External id": 1568,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 978, "finished": true, "Ev Idx": 1567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564830423.591, "dur": 39.951,
    "args": {
      "External id": 1569,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 979, "finished": true, "Ev Idx": 1568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564830471.354, "dur": 44.701,
    "args": {
      "External id": 1570,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 980, "finished": true, "Ev Idx": 1569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564830471.901, "dur": 3.936,
    "args": {
      "External id": 1571,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 980, "finished": true, "Ev Idx": 1570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564830473.239, "dur": 1.475,
    "args": {
      "External id": 1572,"Record function id": 0, "finished": true, "Ev Idx": 1571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830473.886, "dur": 0.688,
    "args": {
      "External id": 1573,"Record function id": 0, "finished": true, "Ev Idx": 1572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564830476.193, "dur": 39.760,
    "args": {
      "External id": 1574,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 981, "finished": true, "Ev Idx": 1573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564830476.668, "dur": 2.099,
    "args": {
      "External id": 1575,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 981, "finished": true, "Ev Idx": 1574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830476.995, "dur": 1.650,
    "args": {
      "External id": 1576,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 981, "finished": true, "Ev Idx": 1575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564830479.133, "dur": 34.922,
    "args": {
      "External id": 1577,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 982, "finished": true, "Ev Idx": 1576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564830490.493, "dur": 1.113,
    "args": {
      "External id": 1578,"Record function id": 0, "finished": true, "Ev Idx": 1577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564830514.624, "dur": 1.017,
    "args": {
      "External id": 1579,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 983, "finished": true, "Ev Idx": 1578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564830518.889, "dur": 8.355,
    "args": {
      "External id": 1580,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 984, "finished": true, "Ev Idx": 1579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564830548.232, "dur": 0.239,
    "args": {
      "External id": 1581,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 985, "finished": true, "Ev Idx": 1580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564830550.478, "dur": 9.272,
    "args": {
      "External id": 1582,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 985, "finished": true, "Ev Idx": 1581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564830551.832, "dur": 0.454,
    "args": {
      "External id": 1583,"Record function id": 0, "finished": true, "Ev Idx": 1582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564830552.587, "dur": 0.068,
    "args": {
      "External id": 1584,"Record function id": 0, "finished": true, "Ev Idx": 1583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564830562.476, "dur": 11.929,
    "args": {
      "External id": 1585,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 986, "finished": true, "Ev Idx": 1584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564830578.234, "dur": 9.696,
    "args": {
      "External id": 1586,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 987, "finished": true, "Ev Idx": 1585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564830589.578, "dur": 9.603,
    "args": {
      "External id": 1587,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 988, "finished": true, "Ev Idx": 1586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564830600.929, "dur": 8.702,
    "args": {
      "External id": 1588,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 989, "finished": true, "Ev Idx": 1587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564830612.369, "dur": 0.085,
    "args": {
      "External id": 1589,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 990, "finished": true, "Ev Idx": 1588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564830613.418, "dur": 8.860,
    "args": {
      "External id": 1590,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 990, "finished": true, "Ev Idx": 1589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564830640.618, "dur": 39.704,
    "args": {
      "External id": 1591,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 991, "finished": true, "Ev Idx": 1590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564830641.237, "dur": 4.070,
    "args": {
      "External id": 1592,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 991, "finished": true, "Ev Idx": 1591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564830643.131, "dur": 1.288,
    "args": {
      "External id": 1593,"Record function id": 0, "finished": true, "Ev Idx": 1592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830643.669, "dur": 0.618,
    "args": {
      "External id": 1594,"Record function id": 0, "finished": true, "Ev Idx": 1593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564830645.958, "dur": 34.236,
    "args": {
      "External id": 1595,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 992, "finished": true, "Ev Idx": 1594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564830646.422, "dur": 1.797,
    "args": {
      "External id": 1596,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 992, "finished": true, "Ev Idx": 1595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830646.986, "dur": 1.103,
    "args": {
      "External id": 1597,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 992, "finished": true, "Ev Idx": 1596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564830648.526, "dur": 29.632,
    "args": {
      "External id": 1598,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 993, "finished": true, "Ev Idx": 1597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564830662.631, "dur": 1.153,
    "args": {
      "External id": 1599,"Record function id": 0, "finished": true, "Ev Idx": 1598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564830678.764, "dur": 1.110,
    "args": {
      "External id": 1600,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 994, "finished": true, "Ev Idx": 1599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830683.595, "dur": 1.411,
    "args": {
      "External id": 1601,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 995, "finished": true, "Ev Idx": 1600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564830686.994, "dur": 2.467,
    "args": {
      "External id": 1602,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 996, "finished": true, "Ev Idx": 1601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830688.100, "dur": 0.509,
    "args": {
      "External id": 1603,"Record function id": 0, "finished": true, "Ev Idx": 1602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564830697.564, "dur": 42.145,
    "args": {
      "External id": 1604,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 997, "finished": true, "Ev Idx": 1603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564830697.881, "dur": 2.175,
    "args": {
      "External id": 1605,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 997, "finished": true, "Ev Idx": 1604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564830698.762, "dur": 0.840,
    "args": {
      "External id": 1606,"Record function id": 0, "finished": true, "Ev Idx": 1605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830699.232, "dur": 0.276,
    "args": {
      "External id": 1607,"Record function id": 0, "finished": true, "Ev Idx": 1606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564830700.391, "dur": 39.215,
    "args": {
      "External id": 1608,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 998, "finished": true, "Ev Idx": 1607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564830700.919, "dur": 1.230,
    "args": {
      "External id": 1609,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 998, "finished": true, "Ev Idx": 1608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830701.164, "dur": 0.874,
    "args": {
      "External id": 1610,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 998, "finished": true, "Ev Idx": 1609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564830702.451, "dur": 35.315,
    "args": {
      "External id": 1611,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 999, "finished": true, "Ev Idx": 1610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564830714.118, "dur": 1.081,
    "args": {
      "External id": 1612,"Record function id": 0, "finished": true, "Ev Idx": 1611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564830738.374, "dur": 0.928,
    "args": {
      "External id": 1613,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1000, "finished": true, "Ev Idx": 1612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830741.971, "dur": 1.162,
    "args": {
      "External id": 1614,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1001, "finished": true, "Ev Idx": 1613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564830744.542, "dur": 1.579,
    "args": {
      "External id": 1615,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1002, "finished": true, "Ev Idx": 1614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830745.276, "dur": 0.384,
    "args": {
      "External id": 1616,"Record function id": 0, "finished": true, "Ev Idx": 1615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564830753.339, "dur": 40.852,
    "args": {
      "External id": 1617,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1003, "finished": true, "Ev Idx": 1616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564830753.625, "dur": 2.090,
    "args": {
      "External id": 1618,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1003, "finished": true, "Ev Idx": 1617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564830754.469, "dur": 0.814,
    "args": {
      "External id": 1619,"Record function id": 0, "finished": true, "Ev Idx": 1618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830754.930, "dur": 0.248,
    "args": {
      "External id": 1620,"Record function id": 0, "finished": true, "Ev Idx": 1619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564830756.015, "dur": 38.075,
    "args": {
      "External id": 1621,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1004, "finished": true, "Ev Idx": 1620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564830756.555, "dur": 1.210,
    "args": {
      "External id": 1622,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1004, "finished": true, "Ev Idx": 1621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830756.831, "dur": 0.829,
    "args": {
      "External id": 1623,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1004, "finished": true, "Ev Idx": 1622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564830758.228, "dur": 34.167,
    "args": {
      "External id": 1624,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1005, "finished": true, "Ev Idx": 1623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564830771.561, "dur": 1.077,
    "args": {
      "External id": 1625,"Record function id": 0, "finished": true, "Ev Idx": 1624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564830792.906, "dur": 0.940,
    "args": {
      "External id": 1626,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1006, "finished": true, "Ev Idx": 1625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564830796.432, "dur": 1.097,
    "args": {
      "External id": 1627,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1007, "finished": true, "Ev Idx": 1626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564830798.921, "dur": 1.888,
    "args": {
      "External id": 1628,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1008, "finished": true, "Ev Idx": 1627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830800.004, "dur": 0.361,
    "args": {
      "External id": 1629,"Record function id": 0, "finished": true, "Ev Idx": 1628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564830803.811, "dur": 2.242,
    "args": {
      "External id": 1630,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1009, "finished": true, "Ev Idx": 1629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830805.219, "dur": 0.298,
    "args": {
      "External id": 1631,"Record function id": 0, "finished": true, "Ev Idx": 1630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564830807.203, "dur": 1.235,
    "args": {
      "External id": 1632,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1009, "finished": true, "Ev Idx": 1631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830807.686, "dur": 0.267,
    "args": {
      "External id": 1633,"Record function id": 0, "finished": true, "Ev Idx": 1632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564830809.424, "dur": 9.300,
    "args": {
      "External id": 1634,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1009, "finished": true, "Ev Idx": 1633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564830825.157, "dur": 5.518,
    "args": {
      "External id": 1635,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1010, "finished": true, "Ev Idx": 1634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830827.921, "dur": 0.808,
    "args": {
      "External id": 1636,"Record function id": 0, "finished": true, "Ev Idx": 1635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564830832.955, "dur": 2.469,
    "args": {
      "External id": 1637,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1011, "finished": true, "Ev Idx": 1636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830834.460, "dur": 0.364,
    "args": {
      "External id": 1638,"Record function id": 0, "finished": true, "Ev Idx": 1637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564830837.457, "dur": 9.138,
    "args": {
      "External id": 1639,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1012, "finished": true, "Ev Idx": 1638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564830851.146, "dur": 18.306,
    "args": {
      "External id": 1640,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1013, "finished": true, "Ev Idx": 1639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564830872.473, "dur": 7.885,
    "args": {
      "External id": 1641,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1014, "finished": true, "Ev Idx": 1640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564830881.885, "dur": 8.479,
    "args": {
      "External id": 1642,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1015, "finished": true, "Ev Idx": 1641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564830892.347, "dur": 8.971,
    "args": {
      "External id": 1643,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1016, "finished": true, "Ev Idx": 1642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564830903.678, "dur": 2.136,
    "args": {
      "External id": 1644,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1017, "finished": true, "Ev Idx": 1643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830904.685, "dur": 0.472,
    "args": {
      "External id": 1645,"Record function id": 0, "finished": true, "Ev Idx": 1644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564830907.372, "dur": 1.722,
    "args": {
      "External id": 1646,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1018, "finished": true, "Ev Idx": 1645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830908.215, "dur": 0.352,
    "args": {
      "External id": 1647,"Record function id": 0, "finished": true, "Ev Idx": 1646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564830910.070, "dur": 7.187,
    "args": {
      "External id": 1648,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1019, "finished": true, "Ev Idx": 1647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564830919.264, "dur": 10.006,
    "args": {
      "External id": 1649,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1020, "finished": true, "Ev Idx": 1648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564830931.492, "dur": 6.553,
    "args": {
      "External id": 1650,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1021, "finished": true, "Ev Idx": 1649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564830939.554, "dur": 9.679,
    "args": {
      "External id": 1651,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1022, "finished": true, "Ev Idx": 1650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564830973.322, "dur": 2.459,
    "args": {
      "External id": 1652,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1023, "finished": true, "Ev Idx": 1651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830974.530, "dur": 0.524,
    "args": {
      "External id": 1653,"Record function id": 0, "finished": true, "Ev Idx": 1652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564830976.464, "dur": 2.727,
    "args": {
      "External id": 1654,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1024, "finished": true, "Ev Idx": 1653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830978.284, "dur": 0.406,
    "args": {
      "External id": 1655,"Record function id": 0, "finished": true, "Ev Idx": 1654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564830979.932, "dur": 3.292,
    "args": {
      "External id": 1656,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1025, "finished": true, "Ev Idx": 1655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830981.881, "dur": 0.715,
    "args": {
      "External id": 1657,"Record function id": 0, "finished": true, "Ev Idx": 1656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564830983.875, "dur": 1.710,
    "args": {
      "External id": 1658,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1026, "finished": true, "Ev Idx": 1657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830984.678, "dur": 0.357,
    "args": {
      "External id": 1659,"Record function id": 0, "finished": true, "Ev Idx": 1658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564830986.229, "dur": 1.954,
    "args": {
      "External id": 1660,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1027, "finished": true, "Ev Idx": 1659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830987.249, "dur": 0.434,
    "args": {
      "External id": 1661,"Record function id": 0, "finished": true, "Ev Idx": 1660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564830991.196, "dur": 3.669,
    "args": {
      "External id": 1662,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1028, "finished": true, "Ev Idx": 1661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564830993.739, "dur": 0.365,
    "args": {
      "External id": 1663,"Record function id": 0, "finished": true, "Ev Idx": 1662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564830997.357, "dur": 20.036,
    "args": {
      "External id": 1664,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1029, "finished": true, "Ev Idx": 1663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564830999.970, "dur": 15.436,
    "args": {
      "External id": 1665,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1029, "finished": true, "Ev Idx": 1664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564831001.826, "dur": 2.576,
    "args": {
      "External id": 1666,"Record function id": 0, "finished": true, "Ev Idx": 1665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831002.622, "dur": 1.561,
    "args": {
      "External id": 1667,"Record function id": 0, "finished": true, "Ev Idx": 1666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564831005.034, "dur": 9.756,
    "args": {
      "External id": 1668,"Record function id": 0, "finished": true, "Ev Idx": 1667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831015.959, "dur": 1.189,
    "args": {
      "External id": 1669,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1030, "finished": true, "Ev Idx": 1668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564831020.590, "dur": 2.991,
    "args": {
      "External id": 1670,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1031, "finished": true, "Ev Idx": 1669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831022.653, "dur": 0.361,
    "args": {
      "External id": 1671,"Record function id": 0, "finished": true, "Ev Idx": 1670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564831024.170, "dur": 2.064,
    "args": {
      "External id": 1672,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1032, "finished": true, "Ev Idx": 1671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831025.501, "dur": 0.288,
    "args": {
      "External id": 1673,"Record function id": 0, "finished": true, "Ev Idx": 1672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564831026.757, "dur": 2.340,
    "args": {
      "External id": 1674,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1033, "finished": true, "Ev Idx": 1673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831028.223, "dur": 0.435,
    "args": {
      "External id": 1675,"Record function id": 0, "finished": true, "Ev Idx": 1674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564831029.800, "dur": 2.254,
    "args": {
      "External id": 1676,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1034, "finished": true, "Ev Idx": 1675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831031.257, "dur": 0.312,
    "args": {
      "External id": 1677,"Record function id": 0, "finished": true, "Ev Idx": 1676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564831032.813, "dur": 1.601,
    "args": {
      "External id": 1678,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1035, "finished": true, "Ev Idx": 1677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831033.668, "dur": 0.275,
    "args": {
      "External id": 1679,"Record function id": 0, "finished": true, "Ev Idx": 1678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564831036.011, "dur": 1.564,
    "args": {
      "External id": 1680,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1036, "finished": true, "Ev Idx": 1679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831036.838, "dur": 0.311,
    "args": {
      "External id": 1681,"Record function id": 0, "finished": true, "Ev Idx": 1680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831038.967, "dur": 12.755,
    "args": {
      "External id": 1682,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1037, "finished": true, "Ev Idx": 1681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564831039.890, "dur": 9.032,
    "args": {
      "External id": 1683,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1037, "finished": true, "Ev Idx": 1682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564831040.704, "dur": 1.599,
    "args": {
      "External id": 1684,"Record function id": 0, "finished": true, "Ev Idx": 1683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831041.059, "dur": 1.121,
    "args": {
      "External id": 1685,"Record function id": 0, "finished": true, "Ev Idx": 1684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564831042.627, "dur": 5.898,
    "args": {
      "External id": 1686,"Record function id": 0, "finished": true, "Ev Idx": 1685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831049.311, "dur": 2.247,
    "args": {
      "External id": 1687,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1038, "finished": true, "Ev Idx": 1686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831053.809, "dur": 2.624,
    "args": {
      "External id": 1688,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1039, "finished": true, "Ev Idx": 1687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831055.568, "dur": 0.379,
    "args": {
      "External id": 1689,"Record function id": 0, "finished": true, "Ev Idx": 1688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564831058.035, "dur": 44.711,
    "args": {
      "External id": 1690,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1040, "finished": true, "Ev Idx": 1689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564831058.637, "dur": 1.538,
    "args": {
      "External id": 1691,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1040, "finished": true, "Ev Idx": 1690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831059.390, "dur": 0.305,
    "args": {
      "External id": 1692,"Record function id": 0, "finished": true, "Ev Idx": 1691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831060.554, "dur": 4.485,
    "args": {
      "External id": 1693,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1041, "finished": true, "Ev Idx": 1692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564831062.378, "dur": 2.517,
    "args": {
      "External id": 1694,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1041, "finished": true, "Ev Idx": 1693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564831065.520, "dur": 1.596,
    "args": {
      "External id": 1695,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1042, "finished": true, "Ev Idx": 1694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831066.415, "dur": 0.327,
    "args": {
      "External id": 1696,"Record function id": 0, "finished": true, "Ev Idx": 1695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831067.422, "dur": 1.929,
    "args": {
      "External id": 1697,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1043, "finished": true, "Ev Idx": 1696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564831068.224, "dur": 1.009,
    "args": {
      "External id": 1698,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1043, "finished": true, "Ev Idx": 1697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564831069.971, "dur": 30.736,
    "args": {
      "External id": 1699,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1044, "finished": true, "Ev Idx": 1698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831084.042, "dur": 1.078,
    "args": {
      "External id": 1700,"Record function id": 0, "finished": true, "Ev Idx": 1699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831101.281, "dur": 1.109,
    "args": {
      "External id": 1701,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1045, "finished": true, "Ev Idx": 1700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564831106.095, "dur": 10.367,
    "args": {
      "External id": 1702,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1046, "finished": true, "Ev Idx": 1701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564831119.549, "dur": 2.236,
    "args": {
      "External id": 1703,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1047, "finished": true, "Ev Idx": 1702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831120.484, "dur": 0.642,
    "args": {
      "External id": 1704,"Record function id": 0, "finished": true, "Ev Idx": 1703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564831122.475, "dur": 1.616,
    "args": {
      "External id": 1705,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1047, "finished": true, "Ev Idx": 1704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831123.242, "dur": 0.418,
    "args": {
      "External id": 1706,"Record function id": 0, "finished": true, "Ev Idx": 1705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564831124.735, "dur": 1.335,
    "args": {
      "External id": 1707,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1047, "finished": true, "Ev Idx": 1706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831125.357, "dur": 0.395,
    "args": {
      "External id": 1708,"Record function id": 0, "finished": true, "Ev Idx": 1707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564831127.770, "dur": 7.448,
    "args": {
      "External id": 1709,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1047, "finished": true, "Ev Idx": 1708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564831140.856, "dur": 12.793,
    "args": {
      "External id": 1710,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1048, "finished": true, "Ev Idx": 1709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564831141.805, "dur": 0.371,
    "args": {
      "External id": 1711,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1048, "finished": true, "Ev Idx": 1710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564831142.863, "dur": 10.502,
    "args": {
      "External id": 1712,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1048, "finished": true, "Ev Idx": 1711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564831156.416, "dur": 0.159,
    "args": {
      "External id": 1713,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1049, "finished": true, "Ev Idx": 1712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564831167.635, "dur": 0.696,
    "args": {
      "External id": 1714,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1049, "finished": true, "Ev Idx": 1713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564831170.152, "dur": 41.282,
    "args": {
      "External id": 1715,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1049, "finished": true, "Ev Idx": 1714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564831170.783, "dur": 2.269,
    "args": {
      "External id": 1716,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1049, "finished": true, "Ev Idx": 1715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831171.958, "dur": 0.476,
    "args": {
      "External id": 1717,"Record function id": 0, "finished": true, "Ev Idx": 1716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831173.433, "dur": 2.979,
    "args": {
      "External id": 1718,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1050, "finished": true, "Ev Idx": 1717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831173.873, "dur": 2.444,
    "args": {
      "External id": 1719,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1050, "finished": true, "Ev Idx": 1718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564831176.965, "dur": 1.823,
    "args": {
      "External id": 1720,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1051, "finished": true, "Ev Idx": 1719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831178.123, "dur": 0.290,
    "args": {
      "External id": 1721,"Record function id": 0, "finished": true, "Ev Idx": 1720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831179.070, "dur": 1.223,
    "args": {
      "External id": 1722,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1052, "finished": true, "Ev Idx": 1721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831179.328, "dur": 0.882,
    "args": {
      "External id": 1723,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1052, "finished": true, "Ev Idx": 1722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564831180.737, "dur": 28.704,
    "args": {
      "External id": 1724,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1053, "finished": true, "Ev Idx": 1723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831191.134, "dur": 1.161,
    "args": {
      "External id": 1725,"Record function id": 0, "finished": true, "Ev Idx": 1724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831209.957, "dur": 1.166,
    "args": {
      "External id": 1726,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1054, "finished": true, "Ev Idx": 1725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831213.476, "dur": 1.759,
    "args": {
      "External id": 1727,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1055, "finished": true, "Ev Idx": 1726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831214.308, "dur": 0.406,
    "args": {
      "External id": 1728,"Record function id": 0, "finished": true, "Ev Idx": 1727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564831217.661, "dur": 12.258,
    "args": {
      "External id": 1729,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1056, "finished": true, "Ev Idx": 1728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564831218.405, "dur": 11.371,
    "args": {
      "External id": 1730,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1056, "finished": true, "Ev Idx": 1729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564831219.508, "dur": 1.948,
    "args": {
      "External id": 1731,"Record function id": 0, "finished": true, "Ev Idx": 1730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831219.934, "dur": 1.388,
    "args": {
      "External id": 1732,"Record function id": 0, "finished": true, "Ev Idx": 1731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564831221.806, "dur": 7.467,
    "args": {
      "External id": 1733,"Record function id": 0, "finished": true, "Ev Idx": 1732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831235.694, "dur": 1.672,
    "args": {
      "External id": 1734,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1057, "finished": true, "Ev Idx": 1733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831235.974, "dur": 1.302,
    "args": {
      "External id": 1735,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1057, "finished": true, "Ev Idx": 1734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564831251.009, "dur": 35.625,
    "args": {
      "External id": 1736,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1058, "finished": true, "Ev Idx": 1735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564831251.576, "dur": 3.293,
    "args": {
      "External id": 1737,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1058, "finished": true, "Ev Idx": 1736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831253.036, "dur": 1.278,
    "args": {
      "External id": 1738,"Record function id": 0, "finished": true, "Ev Idx": 1737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831253.644, "dur": 0.549,
    "args": {
      "External id": 1739,"Record function id": 0, "finished": true, "Ev Idx": 1738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564831255.267, "dur": 31.228,
    "args": {
      "External id": 1740,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1059, "finished": true, "Ev Idx": 1739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831255.834, "dur": 1.478,
    "args": {
      "External id": 1741,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1059, "finished": true, "Ev Idx": 1740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831256.160, "dur": 1.067,
    "args": {
      "External id": 1742,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1059, "finished": true, "Ev Idx": 1741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564831257.734, "dur": 26.744,
    "args": {
      "External id": 1743,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1060, "finished": true, "Ev Idx": 1742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831269.253, "dur": 1.267,
    "args": {
      "External id": 1744,"Record function id": 0, "finished": true, "Ev Idx": 1743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831285.134, "dur": 0.972,
    "args": {
      "External id": 1745,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1061, "finished": true, "Ev Idx": 1744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564831291.681, "dur": 9.411,
    "args": {
      "External id": 1746,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1062, "finished": true, "Ev Idx": 1745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564831310.286, "dur": 0.255,
    "args": {
      "External id": 1747,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1063, "finished": true, "Ev Idx": 1746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564831312.737, "dur": 11.831,
    "args": {
      "External id": 1748,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1063, "finished": true, "Ev Idx": 1747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564831314.460, "dur": 0.383,
    "args": {
      "External id": 1749,"Record function id": 0, "finished": true, "Ev Idx": 1748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564831315.439, "dur": 0.112,
    "args": {
      "External id": 1750,"Record function id": 0, "finished": true, "Ev Idx": 1749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564831327.518, "dur": 11.773,
    "args": {
      "External id": 1751,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1064, "finished": true, "Ev Idx": 1750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564831342.869, "dur": 10.205,
    "args": {
      "External id": 1752,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1065, "finished": true, "Ev Idx": 1751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564831354.881, "dur": 9.683,
    "args": {
      "External id": 1753,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1066, "finished": true, "Ev Idx": 1752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564831366.414, "dur": 7.578,
    "args": {
      "External id": 1754,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1067, "finished": true, "Ev Idx": 1753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564831376.787, "dur": 0.292,
    "args": {
      "External id": 1755,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1068, "finished": true, "Ev Idx": 1754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564831378.090, "dur": 8.159,
    "args": {
      "External id": 1756,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1068, "finished": true, "Ev Idx": 1755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564831400.151, "dur": 123.268,
    "args": {
      "External id": 1757,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1069, "finished": true, "Ev Idx": 1756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564831400.920, "dur": 3.406,
    "args": {
      "External id": 1758,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1069, "finished": true, "Ev Idx": 1757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831402.256, "dur": 1.489,
    "args": {
      "External id": 1759,"Record function id": 0, "finished": true, "Ev Idx": 1758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831403.126, "dur": 0.467,
    "args": {
      "External id": 1760,"Record function id": 0, "finished": true, "Ev Idx": 1759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564831404.878, "dur": 118.399,
    "args": {
      "External id": 1761,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1070, "finished": true, "Ev Idx": 1760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831405.320, "dur": 2.749,
    "args": {
      "External id": 1762,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1070, "finished": true, "Ev Idx": 1761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831405.607, "dur": 2.365,
    "args": {
      "External id": 1763,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1070, "finished": true, "Ev Idx": 1762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564831408.398, "dur": 112.438,
    "args": {
      "External id": 1764,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1071, "finished": true, "Ev Idx": 1763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831477.859, "dur": 13.890,
    "args": {
      "External id": 1765,"Record function id": 0, "finished": true, "Ev Idx": 1764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831521.635, "dur": 1.242,
    "args": {
      "External id": 1766,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1072, "finished": true, "Ev Idx": 1765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564831534.522, "dur": 11.661,
    "args": {
      "External id": 1767,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1073, "finished": true, "Ev Idx": 1766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564831555.112, "dur": 79.207,
    "args": {
      "External id": 1768,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1074, "finished": true, "Ev Idx": 1767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564831555.637, "dur": 3.303,
    "args": {
      "External id": 1769,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1074, "finished": true, "Ev Idx": 1768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831556.871, "dur": 1.243,
    "args": {
      "External id": 1770,"Record function id": 0, "finished": true, "Ev Idx": 1769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831557.432, "dur": 0.548,
    "args": {
      "External id": 1771,"Record function id": 0, "finished": true, "Ev Idx": 1770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564831559.348, "dur": 74.844,
    "args": {
      "External id": 1772,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1075, "finished": true, "Ev Idx": 1771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831559.780, "dur": 1.919,
    "args": {
      "External id": 1773,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1075, "finished": true, "Ev Idx": 1772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831560.200, "dur": 1.390,
    "args": {
      "External id": 1774,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1075, "finished": true, "Ev Idx": 1773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564831562.065, "dur": 69.729,
    "args": {
      "External id": 1775,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1076, "finished": true, "Ev Idx": 1774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831606.283, "dur": 1.206,
    "args": {
      "External id": 1776,"Record function id": 0, "finished": true, "Ev Idx": 1775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831632.444, "dur": 1.423,
    "args": {
      "External id": 1777,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1077, "finished": true, "Ev Idx": 1776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564831636.802, "dur": 9.540,
    "args": {
      "External id": 1778,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1078, "finished": true, "Ev Idx": 1777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564831653.676, "dur": 41.491,
    "args": {
      "External id": 1779,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1079, "finished": true, "Ev Idx": 1778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564831654.481, "dur": 2.683,
    "args": {
      "External id": 1780,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1079, "finished": true, "Ev Idx": 1779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831655.494, "dur": 1.083,
    "args": {
      "External id": 1781,"Record function id": 0, "finished": true, "Ev Idx": 1780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831655.987, "dur": 0.458,
    "args": {
      "External id": 1782,"Record function id": 0, "finished": true, "Ev Idx": 1781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564831657.478, "dur": 37.587,
    "args": {
      "External id": 1783,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1080, "finished": true, "Ev Idx": 1782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831658.019, "dur": 1.383,
    "args": {
      "External id": 1784,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1080, "finished": true, "Ev Idx": 1783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831658.322, "dur": 0.995,
    "args": {
      "External id": 1785,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1080, "finished": true, "Ev Idx": 1784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564831659.760, "dur": 33.463,
    "args": {
      "External id": 1786,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1081, "finished": true, "Ev Idx": 1785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831671.434, "dur": 1.219,
    "args": {
      "External id": 1787,"Record function id": 0, "finished": true, "Ev Idx": 1786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831693.792, "dur": 0.939,
    "args": {
      "External id": 1788,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1082, "finished": true, "Ev Idx": 1787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564831697.937, "dur": 9.533,
    "args": {
      "External id": 1789,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1083, "finished": true, "Ev Idx": 1788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564831728.876, "dur": 0.257,
    "args": {
      "External id": 1790,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1084, "finished": true, "Ev Idx": 1789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564831731.163, "dur": 11.564,
    "args": {
      "External id": 1791,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1084, "finished": true, "Ev Idx": 1790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564831732.598, "dur": 0.591,
    "args": {
      "External id": 1792,"Record function id": 0, "finished": true, "Ev Idx": 1791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564831733.544, "dur": 0.104,
    "args": {
      "External id": 1793,"Record function id": 0, "finished": true, "Ev Idx": 1792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564831745.380, "dur": 12.826,
    "args": {
      "External id": 1794,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1085, "finished": true, "Ev Idx": 1793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564831761.971, "dur": 8.413,
    "args": {
      "External id": 1795,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1086, "finished": true, "Ev Idx": 1794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564831772.132, "dur": 8.383,
    "args": {
      "External id": 1796,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1087, "finished": true, "Ev Idx": 1795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564831782.199, "dur": 9.232,
    "args": {
      "External id": 1797,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1088, "finished": true, "Ev Idx": 1796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564831794.019, "dur": 0.109,
    "args": {
      "External id": 1798,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1089, "finished": true, "Ev Idx": 1797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564831794.985, "dur": 9.367,
    "args": {
      "External id": 1799,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1089, "finished": true, "Ev Idx": 1798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564831822.542, "dur": 38.971,
    "args": {
      "External id": 1800,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1090, "finished": true, "Ev Idx": 1799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564831822.988, "dur": 4.054,
    "args": {
      "External id": 1801,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1090, "finished": true, "Ev Idx": 1800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831824.858, "dur": 1.216,
    "args": {
      "External id": 1802,"Record function id": 0, "finished": true, "Ev Idx": 1801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831825.505, "dur": 0.438,
    "args": {
      "External id": 1803,"Record function id": 0, "finished": true, "Ev Idx": 1802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564831827.349, "dur": 34.041,
    "args": {
      "External id": 1804,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1091, "finished": true, "Ev Idx": 1803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831827.904, "dur": 1.633,
    "args": {
      "External id": 1805,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1091, "finished": true, "Ev Idx": 1804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831828.208, "dur": 1.212,
    "args": {
      "External id": 1806,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1091, "finished": true, "Ev Idx": 1805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564831829.909, "dur": 29.350,
    "args": {
      "External id": 1807,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1092, "finished": true, "Ev Idx": 1806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831841.545, "dur": 1.142,
    "args": {
      "External id": 1808,"Record function id": 0, "finished": true, "Ev Idx": 1807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831859.912, "dur": 1.158,
    "args": {
      "External id": 1809,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1093, "finished": true, "Ev Idx": 1808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831864.999, "dur": 1.467,
    "args": {
      "External id": 1810,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1094, "finished": true, "Ev Idx": 1809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831868.400, "dur": 2.629,
    "args": {
      "External id": 1811,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1095, "finished": true, "Ev Idx": 1810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831869.571, "dur": 0.559,
    "args": {
      "External id": 1812,"Record function id": 0, "finished": true, "Ev Idx": 1811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564831878.575, "dur": 42.252,
    "args": {
      "External id": 1813,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1096, "finished": true, "Ev Idx": 1812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564831878.928, "dur": 2.260,
    "args": {
      "External id": 1814,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1096, "finished": true, "Ev Idx": 1813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831879.878, "dur": 0.875,
    "args": {
      "External id": 1815,"Record function id": 0, "finished": true, "Ev Idx": 1814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831880.366, "dur": 0.278,
    "args": {
      "External id": 1816,"Record function id": 0, "finished": true, "Ev Idx": 1815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564831881.512, "dur": 39.212,
    "args": {
      "External id": 1817,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1097, "finished": true, "Ev Idx": 1816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831882.081, "dur": 1.351,
    "args": {
      "External id": 1818,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1097, "finished": true, "Ev Idx": 1817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831882.345, "dur": 0.982,
    "args": {
      "External id": 1819,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1097, "finished": true, "Ev Idx": 1818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564831883.748, "dur": 35.139,
    "args": {
      "External id": 1820,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1098, "finished": true, "Ev Idx": 1819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831896.043, "dur": 1.079,
    "args": {
      "External id": 1821,"Record function id": 0, "finished": true, "Ev Idx": 1820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831919.449, "dur": 0.983,
    "args": {
      "External id": 1822,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1099, "finished": true, "Ev Idx": 1821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831923.111, "dur": 1.211,
    "args": {
      "External id": 1823,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1100, "finished": true, "Ev Idx": 1822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831925.969, "dur": 1.730,
    "args": {
      "External id": 1824,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1101, "finished": true, "Ev Idx": 1823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831926.734, "dur": 0.424,
    "args": {
      "External id": 1825,"Record function id": 0, "finished": true, "Ev Idx": 1824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564831934.224, "dur": 39.432,
    "args": {
      "External id": 1826,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1102, "finished": true, "Ev Idx": 1825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564831934.656, "dur": 2.075,
    "args": {
      "External id": 1827,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1102, "finished": true, "Ev Idx": 1826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831935.419, "dur": 0.901,
    "args": {
      "External id": 1828,"Record function id": 0, "finished": true, "Ev Idx": 1827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831935.881, "dur": 0.305,
    "args": {
      "External id": 1829,"Record function id": 0, "finished": true, "Ev Idx": 1828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564831937.044, "dur": 36.514,
    "args": {
      "External id": 1830,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1103, "finished": true, "Ev Idx": 1829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564831937.449, "dur": 1.459,
    "args": {
      "External id": 1831,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1103, "finished": true, "Ev Idx": 1830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831937.915, "dur": 0.881,
    "args": {
      "External id": 1832,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1103, "finished": true, "Ev Idx": 1831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564831939.230, "dur": 32.607,
    "args": {
      "External id": 1833,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1104, "finished": true, "Ev Idx": 1832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564831950.801, "dur": 1.044,
    "args": {
      "External id": 1834,"Record function id": 0, "finished": true, "Ev Idx": 1833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564831972.333, "dur": 0.929,
    "args": {
      "External id": 1835,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1105, "finished": true, "Ev Idx": 1834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564831976.056, "dur": 1.184,
    "args": {
      "External id": 1836,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1106, "finished": true, "Ev Idx": 1835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564831978.753, "dur": 1.761,
    "args": {
      "External id": 1837,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1107, "finished": true, "Ev Idx": 1836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831979.640, "dur": 0.386,
    "args": {
      "External id": 1838,"Record function id": 0, "finished": true, "Ev Idx": 1837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564831983.374, "dur": 2.161,
    "args": {
      "External id": 1839,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1108, "finished": true, "Ev Idx": 1838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831984.674, "dur": 0.332,
    "args": {
      "External id": 1840,"Record function id": 0, "finished": true, "Ev Idx": 1839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564831986.782, "dur": 1.301,
    "args": {
      "External id": 1841,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1108, "finished": true, "Ev Idx": 1840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564831987.253, "dur": 0.492,
    "args": {
      "External id": 1842,"Record function id": 0, "finished": true, "Ev Idx": 1841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564831989.096, "dur": 49.413,
    "args": {
      "External id": 1843,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1108, "finished": true, "Ev Idx": 1842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832044.896, "dur": 4.825,
    "args": {
      "External id": 1844,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1109, "finished": true, "Ev Idx": 1843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832047.472, "dur": 0.601,
    "args": {
      "External id": 1845,"Record function id": 0, "finished": true, "Ev Idx": 1844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832052.177, "dur": 2.694,
    "args": {
      "External id": 1846,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1110, "finished": true, "Ev Idx": 1845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832053.801, "dur": 0.463,
    "args": {
      "External id": 1847,"Record function id": 0, "finished": true, "Ev Idx": 1846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564832056.820, "dur": 13.422,
    "args": {
      "External id": 1848,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1111, "finished": true, "Ev Idx": 1847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564832074.770, "dur": 17.095,
    "args": {
      "External id": 1849,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1112, "finished": true, "Ev Idx": 1848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564832095.643, "dur": 9.613,
    "args": {
      "External id": 1850,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1113, "finished": true, "Ev Idx": 1849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564832106.913, "dur": 8.448,
    "args": {
      "External id": 1851,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1114, "finished": true, "Ev Idx": 1850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564832117.591, "dur": 6.720,
    "args": {
      "External id": 1852,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1115, "finished": true, "Ev Idx": 1851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832126.670, "dur": 3.023,
    "args": {
      "External id": 1853,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1116, "finished": true, "Ev Idx": 1852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832128.535, "dur": 0.446,
    "args": {
      "External id": 1854,"Record function id": 0, "finished": true, "Ev Idx": 1853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832131.250, "dur": 2.643,
    "args": {
      "External id": 1855,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1117, "finished": true, "Ev Idx": 1854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832133.028, "dur": 0.319,
    "args": {
      "External id": 1856,"Record function id": 0, "finished": true, "Ev Idx": 1855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564832134.987, "dur": 7.769,
    "args": {
      "External id": 1857,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1118, "finished": true, "Ev Idx": 1856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564832144.880, "dur": 11.824,
    "args": {
      "External id": 1858,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1119, "finished": true, "Ev Idx": 1857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564832159.073, "dur": 7.966,
    "args": {
      "External id": 1859,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1120, "finished": true, "Ev Idx": 1858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564832168.332, "dur": 6.564,
    "args": {
      "External id": 1860,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1121, "finished": true, "Ev Idx": 1859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832199.581, "dur": 2.650,
    "args": {
      "External id": 1861,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1122, "finished": true, "Ev Idx": 1860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832200.909, "dur": 0.549,
    "args": {
      "External id": 1862,"Record function id": 0, "finished": true, "Ev Idx": 1861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832202.872, "dur": 2.831,
    "args": {
      "External id": 1863,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1123, "finished": true, "Ev Idx": 1862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832204.798, "dur": 0.459,
    "args": {
      "External id": 1864,"Record function id": 0, "finished": true, "Ev Idx": 1863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564832206.587, "dur": 3.298,
    "args": {
      "External id": 1865,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1124, "finished": true, "Ev Idx": 1864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832208.490, "dur": 0.706,
    "args": {
      "External id": 1866,"Record function id": 0, "finished": true, "Ev Idx": 1865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832210.663, "dur": 2.141,
    "args": {
      "External id": 1867,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1125, "finished": true, "Ev Idx": 1866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832211.752, "dur": 0.569,
    "args": {
      "External id": 1868,"Record function id": 0, "finished": true, "Ev Idx": 1867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832213.412, "dur": 1.539,
    "args": {
      "External id": 1869,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1126, "finished": true, "Ev Idx": 1868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832214.212, "dur": 0.279,
    "args": {
      "External id": 1870,"Record function id": 0, "finished": true, "Ev Idx": 1869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564832217.981, "dur": 2.877,
    "args": {
      "External id": 1871,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1127, "finished": true, "Ev Idx": 1870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832219.640, "dur": 0.497,
    "args": {
      "External id": 1872,"Record function id": 0, "finished": true, "Ev Idx": 1871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832223.367, "dur": 20.419,
    "args": {
      "External id": 1873,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1128, "finished": true, "Ev Idx": 1872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564832226.053, "dur": 15.055,
    "args": {
      "External id": 1874,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1128, "finished": true, "Ev Idx": 1873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564832227.452, "dur": 3.206,
    "args": {
      "External id": 1875,"Record function id": 0, "finished": true, "Ev Idx": 1874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564832228.460, "dur": 1.943,
    "args": {
      "External id": 1876,"Record function id": 0, "finished": true, "Ev Idx": 1875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564832231.323, "dur": 9.242,
    "args": {
      "External id": 1877,"Record function id": 0, "finished": true, "Ev Idx": 1876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564832241.868, "dur": 1.647,
    "args": {
      "External id": 1878,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1129, "finished": true, "Ev Idx": 1877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832246.998, "dur": 2.178,
    "args": {
      "External id": 1879,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1130, "finished": true, "Ev Idx": 1878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832248.224, "dur": 0.399,
    "args": {
      "External id": 1880,"Record function id": 0, "finished": true, "Ev Idx": 1879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832249.742, "dur": 1.437,
    "args": {
      "External id": 1881,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1131, "finished": true, "Ev Idx": 1880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832250.456, "dur": 0.287,
    "args": {
      "External id": 1882,"Record function id": 0, "finished": true, "Ev Idx": 1881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564832251.704, "dur": 1.683,
    "args": {
      "External id": 1883,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1132, "finished": true, "Ev Idx": 1882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832252.625, "dur": 0.315,
    "args": {
      "External id": 1884,"Record function id": 0, "finished": true, "Ev Idx": 1883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832254.041, "dur": 1.464,
    "args": {
      "External id": 1885,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1133, "finished": true, "Ev Idx": 1884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832254.728, "dur": 0.310,
    "args": {
      "External id": 1886,"Record function id": 0, "finished": true, "Ev Idx": 1885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832256.072, "dur": 2.662,
    "args": {
      "External id": 1887,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1134, "finished": true, "Ev Idx": 1886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832257.919, "dur": 0.384,
    "args": {
      "External id": 1888,"Record function id": 0, "finished": true, "Ev Idx": 1887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564832260.380, "dur": 2.221,
    "args": {
      "External id": 1889,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1135, "finished": true, "Ev Idx": 1888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832261.834, "dur": 0.334,
    "args": {
      "External id": 1890,"Record function id": 0, "finished": true, "Ev Idx": 1889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832264.059, "dur": 11.618,
    "args": {
      "External id": 1891,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1136, "finished": true, "Ev Idx": 1890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564832265.080, "dur": 9.103,
    "args": {
      "External id": 1892,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1136, "finished": true, "Ev Idx": 1891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564832265.883, "dur": 2.025,
    "args": {
      "External id": 1893,"Record function id": 0, "finished": true, "Ev Idx": 1892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564832266.353, "dur": 1.444,
    "args": {
      "External id": 1894,"Record function id": 0, "finished": true, "Ev Idx": 1893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564832268.246, "dur": 5.597,
    "args": {
      "External id": 1895,"Record function id": 0, "finished": true, "Ev Idx": 1894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564832274.535, "dur": 0.937,
    "args": {
      "External id": 1896,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1137, "finished": true, "Ev Idx": 1895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564832277.889, "dur": 1.792,
    "args": {
      "External id": 1897,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1138, "finished": true, "Ev Idx": 1896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832278.762, "dur": 0.395,
    "args": {
      "External id": 1898,"Record function id": 0, "finished": true, "Ev Idx": 1897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564832281.345, "dur": 47.493,
    "args": {
      "External id": 1899,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1139, "finished": true, "Ev Idx": 1898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564832282.003, "dur": 2.275,
    "args": {
      "External id": 1900,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1139, "finished": true, "Ev Idx": 1899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832283.467, "dur": 0.341,
    "args": {
      "External id": 1901,"Record function id": 0, "finished": true, "Ev Idx": 1900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832284.622, "dur": 4.524,
    "args": {
      "External id": 1902,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1140, "finished": true, "Ev Idx": 1901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564832286.405, "dur": 2.538,
    "args": {
      "External id": 1903,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1140, "finished": true, "Ev Idx": 1902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564832289.628, "dur": 1.542,
    "args": {
      "External id": 1904,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1141, "finished": true, "Ev Idx": 1903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832290.531, "dur": 0.283,
    "args": {
      "External id": 1905,"Record function id": 0, "finished": true, "Ev Idx": 1904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832291.451, "dur": 1.782,
    "args": {
      "External id": 1906,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1142, "finished": true, "Ev Idx": 1905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564832292.308, "dur": 0.810,
    "args": {
      "External id": 1907,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1142, "finished": true, "Ev Idx": 1906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564832293.854, "dur": 33.027,
    "args": {
      "External id": 1908,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1143, "finished": true, "Ev Idx": 1907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564832306.819, "dur": 1.007,
    "args": {
      "External id": 1909,"Record function id": 0, "finished": true, "Ev Idx": 1908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564832327.482, "dur": 1.050,
    "args": {
      "External id": 1910,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1144, "finished": true, "Ev Idx": 1909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564832332.383, "dur": 10.124,
    "args": {
      "External id": 1911,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1145, "finished": true, "Ev Idx": 1910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832345.482, "dur": 1.801,
    "args": {
      "External id": 1912,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1146, "finished": true, "Ev Idx": 1911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832346.217, "dur": 0.468,
    "args": {
      "External id": 1913,"Record function id": 0, "finished": true, "Ev Idx": 1912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832347.882, "dur": 1.333,
    "args": {
      "External id": 1914,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1146, "finished": true, "Ev Idx": 1913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832348.466, "dur": 0.351,
    "args": {
      "External id": 1915,"Record function id": 0, "finished": true, "Ev Idx": 1914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564832349.835, "dur": 1.417,
    "args": {
      "External id": 1916,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1146, "finished": true, "Ev Idx": 1915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832350.502, "dur": 0.431,
    "args": {
      "External id": 1917,"Record function id": 0, "finished": true, "Ev Idx": 1916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564832353.135, "dur": 8.054,
    "args": {
      "External id": 1918,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1146, "finished": true, "Ev Idx": 1917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564832366.244, "dur": 13.851,
    "args": {
      "External id": 1919,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1147, "finished": true, "Ev Idx": 1918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564832367.264, "dur": 0.404,
    "args": {
      "External id": 1920,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1147, "finished": true, "Ev Idx": 1919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564832368.319, "dur": 11.503,
    "args": {
      "External id": 1921,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1147, "finished": true, "Ev Idx": 1920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564832382.987, "dur": 0.176,
    "args": {
      "External id": 1922,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1148, "finished": true, "Ev Idx": 1921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564832393.538, "dur": 0.485,
    "args": {
      "External id": 1923,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1148, "finished": true, "Ev Idx": 1922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564832395.794, "dur": 41.672,
    "args": {
      "External id": 1924,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1148, "finished": true, "Ev Idx": 1923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564832396.326, "dur": 2.525,
    "args": {
      "External id": 1925,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1148, "finished": true, "Ev Idx": 1924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832397.887, "dur": 0.465,
    "args": {
      "External id": 1926,"Record function id": 0, "finished": true, "Ev Idx": 1925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832399.264, "dur": 2.922,
    "args": {
      "External id": 1927,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1149, "finished": true, "Ev Idx": 1926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564832399.574, "dur": 2.492,
    "args": {
      "External id": 1928,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1149, "finished": true, "Ev Idx": 1927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564832402.581, "dur": 1.912,
    "args": {
      "External id": 1929,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1150, "finished": true, "Ev Idx": 1928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832403.769, "dur": 0.354,
    "args": {
      "External id": 1930,"Record function id": 0, "finished": true, "Ev Idx": 1929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832404.759, "dur": 2.060,
    "args": {
      "External id": 1931,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1151, "finished": true, "Ev Idx": 1930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564832405.047, "dur": 1.700,
    "args": {
      "External id": 1932,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1151, "finished": true, "Ev Idx": 1931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564832407.191, "dur": 28.234,
    "args": {
      "External id": 1933,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1152, "finished": true, "Ev Idx": 1932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564832417.564, "dur": 1.091,
    "args": {
      "External id": 1934,"Record function id": 0, "finished": true, "Ev Idx": 1933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564832436.084, "dur": 1.079,
    "args": {
      "External id": 1935,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1153, "finished": true, "Ev Idx": 1934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564832439.429, "dur": 1.775,
    "args": {
      "External id": 1936,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1154, "finished": true, "Ev Idx": 1935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832440.313, "dur": 0.375,
    "args": {
      "External id": 1937,"Record function id": 0, "finished": true, "Ev Idx": 1936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564832443.707, "dur": 11.680,
    "args": {
      "External id": 1938,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1155, "finished": true, "Ev Idx": 1937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564832444.308, "dur": 10.937,
    "args": {
      "External id": 1939,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1155, "finished": true, "Ev Idx": 1938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564832445.371, "dur": 1.873,
    "args": {
      "External id": 1940,"Record function id": 0, "finished": true, "Ev Idx": 1939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564832445.721, "dur": 1.387,
    "args": {
      "External id": 1941,"Record function id": 0, "finished": true, "Ev Idx": 1940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564832447.576, "dur": 7.189,
    "args": {
      "External id": 1942,"Record function id": 0, "finished": true, "Ev Idx": 1941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832460.741, "dur": 1.915,
    "args": {
      "External id": 1943,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1156, "finished": true, "Ev Idx": 1942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564832461.175, "dur": 1.385,
    "args": {
      "External id": 1944,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1156, "finished": true, "Ev Idx": 1943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564832476.549, "dur": 35.388,
    "args": {
      "External id": 1945,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1157, "finished": true, "Ev Idx": 1944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564832477.529, "dur": 3.195,
    "args": {
      "External id": 1946,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1157, "finished": true, "Ev Idx": 1945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564832478.972, "dur": 1.174,
    "args": {
      "External id": 1947,"Record function id": 0, "finished": true, "Ev Idx": 1946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832479.665, "dur": 0.371,
    "args": {
      "External id": 1948,"Record function id": 0, "finished": true, "Ev Idx": 1947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564832481.078, "dur": 30.702,
    "args": {
      "External id": 1949,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1158, "finished": true, "Ev Idx": 1948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832481.647, "dur": 1.867,
    "args": {
      "External id": 1950,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1158, "finished": true, "Ev Idx": 1949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564832481.919, "dur": 1.506,
    "args": {
      "External id": 1951,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1158, "finished": true, "Ev Idx": 1950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564832483.928, "dur": 25.819,
    "args": {
      "External id": 1952,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1159, "finished": true, "Ev Idx": 1951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564832493.735, "dur": 0.981,
    "args": {
      "External id": 1953,"Record function id": 0, "finished": true, "Ev Idx": 1952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564832510.320, "dur": 1.145,
    "args": {
      "External id": 1954,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1160, "finished": true, "Ev Idx": 1953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564832516.440, "dur": 9.416,
    "args": {
      "External id": 1955,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1161, "finished": true, "Ev Idx": 1954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564832534.655, "dur": 0.333,
    "args": {
      "External id": 1956,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1162, "finished": true, "Ev Idx": 1955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564832536.953, "dur": 11.248,
    "args": {
      "External id": 1957,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1162, "finished": true, "Ev Idx": 1956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564832538.986, "dur": 0.615,
    "args": {
      "External id": 1958,"Record function id": 0, "finished": true, "Ev Idx": 1957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564832540.052, "dur": 0.079,
    "args": {
      "External id": 1959,"Record function id": 0, "finished": true, "Ev Idx": 1958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564832550.992, "dur": 14.019,
    "args": {
      "External id": 1960,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1163, "finished": true, "Ev Idx": 1959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564832568.764, "dur": 8.549,
    "args": {
      "External id": 1961,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1164, "finished": true, "Ev Idx": 1960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564832579.127, "dur": 11.668,
    "args": {
      "External id": 1962,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1165, "finished": true, "Ev Idx": 1961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564832592.474, "dur": 7.916,
    "args": {
      "External id": 1963,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1166, "finished": true, "Ev Idx": 1962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564832602.900, "dur": 0.130,
    "args": {
      "External id": 1964,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1167, "finished": true, "Ev Idx": 1963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564832603.934, "dur": 10.249,
    "args": {
      "External id": 1965,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1167, "finished": true, "Ev Idx": 1964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564832628.334, "dur": 50.000,
    "args": {
      "External id": 1966,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1168, "finished": true, "Ev Idx": 1965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564832628.950, "dur": 3.481,
    "args": {
      "External id": 1967,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1168, "finished": true, "Ev Idx": 1966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564832630.481, "dur": 1.383,
    "args": {
      "External id": 1968,"Record function id": 0, "finished": true, "Ev Idx": 1967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832631.371, "dur": 0.379,
    "args": {
      "External id": 1969,"Record function id": 0, "finished": true, "Ev Idx": 1968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564832632.756, "dur": 45.469,
    "args": {
      "External id": 1970,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1169, "finished": true, "Ev Idx": 1969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832633.365, "dur": 2.456,
    "args": {
      "External id": 1971,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1169, "finished": true, "Ev Idx": 1970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564832633.789, "dur": 1.948,
    "args": {
      "External id": 1972,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1169, "finished": true, "Ev Idx": 1971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564832636.311, "dur": 39.889,
    "args": {
      "External id": 1973,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1170, "finished": true, "Ev Idx": 1972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564832648.353, "dur": 1.072,
    "args": {
      "External id": 1974,"Record function id": 0, "finished": true, "Ev Idx": 1973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564832676.978, "dur": 0.960,
    "args": {
      "External id": 1975,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1171, "finished": true, "Ev Idx": 1974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564832686.720, "dur": 62.080,
    "args": {
      "External id": 1976,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1172, "finished": true, "Ev Idx": 1975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564832757.643, "dur": 46.143,
    "args": {
      "External id": 1977,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1173, "finished": true, "Ev Idx": 1976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564832758.261, "dur": 4.344,
    "args": {
      "External id": 1978,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1173, "finished": true, "Ev Idx": 1977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564832760.228, "dur": 1.581,
    "args": {
      "External id": 1979,"Record function id": 0, "finished": true, "Ev Idx": 1978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832760.927, "dur": 0.720,
    "args": {
      "External id": 1980,"Record function id": 0, "finished": true, "Ev Idx": 1979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564832762.943, "dur": 40.737,
    "args": {
      "External id": 1981,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1174, "finished": true, "Ev Idx": 1980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832763.560, "dur": 1.733,
    "args": {
      "External id": 1982,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1174, "finished": true, "Ev Idx": 1981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564832764.117, "dur": 1.064,
    "args": {
      "External id": 1983,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1174, "finished": true, "Ev Idx": 1982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564832765.650, "dur": 36.133,
    "args": {
      "External id": 1984,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1175, "finished": true, "Ev Idx": 1983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564832776.693, "dur": 1.189,
    "args": {
      "External id": 1985,"Record function id": 0, "finished": true, "Ev Idx": 1984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564832802.371, "dur": 0.966,
    "args": {
      "External id": 1986,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1176, "finished": true, "Ev Idx": 1985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564832805.689, "dur": 39.287,
    "args": {
      "External id": 1987,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1177, "finished": true, "Ev Idx": 1986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564832852.390, "dur": 42.654,
    "args": {
      "External id": 1988,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1178, "finished": true, "Ev Idx": 1987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564832852.901, "dur": 3.496,
    "args": {
      "External id": 1989,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1178, "finished": true, "Ev Idx": 1988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564832854.166, "dur": 1.280,
    "args": {
      "External id": 1990,"Record function id": 0, "finished": true, "Ev Idx": 1989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564832854.661, "dur": 0.646,
    "args": {
      "External id": 1991,"Record function id": 0, "finished": true, "Ev Idx": 1990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564832856.711, "dur": 38.231,
    "args": {
      "External id": 1992,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1179, "finished": true, "Ev Idx": 1991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564832857.359, "dur": 2.001,
    "args": {
      "External id": 1993,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1179, "finished": true, "Ev Idx": 1992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564832857.686, "dur": 1.551,
    "args": {
      "External id": 1994,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1179, "finished": true, "Ev Idx": 1993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564832859.726, "dur": 33.196,
    "args": {
      "External id": 1995,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1180, "finished": true, "Ev Idx": 1994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564832869.363, "dur": 1.223,
    "args": {
      "External id": 1996,"Record function id": 0, "finished": true, "Ev Idx": 1995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564832893.542, "dur": 1.079,
    "args": {
      "External id": 1997,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1181, "finished": true, "Ev Idx": 1996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564832897.878, "dur": 8.629,
    "args": {
      "External id": 1998,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1182, "finished": true, "Ev Idx": 1997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564832925.903, "dur": 0.261,
    "args": {
      "External id": 1999,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1183, "finished": true, "Ev Idx": 1998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564832928.060, "dur": 10.227,
    "args": {
      "External id": 2000,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1183, "finished": true, "Ev Idx": 1999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564832929.356, "dur": 0.473,
    "args": {
      "External id": 2001,"Record function id": 0, "finished": true, "Ev Idx": 2000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564832930.221, "dur": 0.073,
    "args": {
      "External id": 2002,"Record function id": 0, "finished": true, "Ev Idx": 2001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564832941.076, "dur": 11.562,
    "args": {
      "External id": 2003,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1184, "finished": true, "Ev Idx": 2002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564832956.368, "dur": 9.738,
    "args": {
      "External id": 2004,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1185, "finished": true, "Ev Idx": 2003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564832967.777, "dur": 8.302,
    "args": {
      "External id": 2005,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1186, "finished": true, "Ev Idx": 2004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564832977.747, "dur": 8.435,
    "args": {
      "External id": 2006,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1187, "finished": true, "Ev Idx": 2005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564832988.468, "dur": 0.120,
    "args": {
      "External id": 2007,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1188, "finished": true, "Ev Idx": 2006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564832989.440, "dur": 7.526,
    "args": {
      "External id": 2008,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1188, "finished": true, "Ev Idx": 2007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564833014.858, "dur": 40.962,
    "args": {
      "External id": 2009,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1189, "finished": true, "Ev Idx": 2008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564833015.284, "dur": 4.545,
    "args": {
      "External id": 2010,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1189, "finished": true, "Ev Idx": 2009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564833017.478, "dur": 1.449,
    "args": {
      "External id": 2011,"Record function id": 0, "finished": true, "Ev Idx": 2010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833018.262, "dur": 0.511,
    "args": {
      "External id": 2012,"Record function id": 0, "finished": true, "Ev Idx": 2011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564833020.318, "dur": 35.386,
    "args": {
      "External id": 2013,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1190, "finished": true, "Ev Idx": 2012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833020.959, "dur": 1.689,
    "args": {
      "External id": 2014,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1190, "finished": true, "Ev Idx": 2013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564833021.452, "dur": 1.094,
    "args": {
      "External id": 2015,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1190, "finished": true, "Ev Idx": 2014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564833023.005, "dur": 30.437,
    "args": {
      "External id": 2016,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1191, "finished": true, "Ev Idx": 2015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564833035.432, "dur": 1.220,
    "args": {
      "External id": 2017,"Record function id": 0, "finished": true, "Ev Idx": 2016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564833054.140, "dur": 1.165,
    "args": {
      "External id": 2018,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1192, "finished": true, "Ev Idx": 2017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564833059.315, "dur": 1.476,
    "args": {
      "External id": 2019,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1193, "finished": true, "Ev Idx": 2018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564833062.585, "dur": 2.553,
    "args": {
      "External id": 2020,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1194, "finished": true, "Ev Idx": 2019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833063.602, "dur": 0.722,
    "args": {
      "External id": 2021,"Record function id": 0, "finished": true, "Ev Idx": 2020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564833222.668, "dur": 50.183,
    "args": {
      "External id": 2022,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1195, "finished": true, "Ev Idx": 2021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564833223.196, "dur": 2.757,
    "args": {
      "External id": 2023,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1195, "finished": true, "Ev Idx": 2022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564833224.325, "dur": 1.063,
    "args": {
      "External id": 2024,"Record function id": 0, "finished": true, "Ev Idx": 2023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833224.909, "dur": 0.383,
    "args": {
      "External id": 2025,"Record function id": 0, "finished": true, "Ev Idx": 2024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564833226.255, "dur": 46.506,
    "args": {
      "External id": 2026,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1196, "finished": true, "Ev Idx": 2025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833226.671, "dur": 1.692,
    "args": {
      "External id": 2027,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1196, "finished": true, "Ev Idx": 2026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564833227.368, "dur": 0.895,
    "args": {
      "External id": 2028,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1196, "finished": true, "Ev Idx": 2027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564833228.849, "dur": 41.776,
    "args": {
      "External id": 2029,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1197, "finished": true, "Ev Idx": 2028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564833242.734, "dur": 1.336,
    "args": {
      "External id": 2030,"Record function id": 0, "finished": true, "Ev Idx": 2029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564833271.272, "dur": 1.175,
    "args": {
      "External id": 2031,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1198, "finished": true, "Ev Idx": 2030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564833276.492, "dur": 1.255,
    "args": {
      "External id": 2032,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1199, "finished": true, "Ev Idx": 2031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564833279.791, "dur": 2.455,
    "args": {
      "External id": 2033,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1200, "finished": true, "Ev Idx": 2032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833281.078, "dur": 0.517,
    "args": {
      "External id": 2034,"Record function id": 0, "finished": true, "Ev Idx": 2033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564833290.935, "dur": 41.688,
    "args": {
      "External id": 2035,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1201, "finished": true, "Ev Idx": 2034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564833291.391, "dur": 2.805,
    "args": {
      "External id": 2036,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1201, "finished": true, "Ev Idx": 2035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564833292.578, "dur": 1.190,
    "args": {
      "External id": 2037,"Record function id": 0, "finished": true, "Ev Idx": 2036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833293.349, "dur": 0.316,
    "args": {
      "External id": 2038,"Record function id": 0, "finished": true, "Ev Idx": 2037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564833294.499, "dur": 38.028,
    "args": {
      "External id": 2039,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1202, "finished": true, "Ev Idx": 2038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833294.926, "dur": 1.374,
    "args": {
      "External id": 2040,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1202, "finished": true, "Ev Idx": 2039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564833295.180, "dur": 1.034,
    "args": {
      "External id": 2041,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1202, "finished": true, "Ev Idx": 2040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564833296.631, "dur": 34.028,
    "args": {
      "External id": 2042,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1203, "finished": true, "Ev Idx": 2041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564833309.078, "dur": 1.112,
    "args": {
      "External id": 2043,"Record function id": 0, "finished": true, "Ev Idx": 2042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564833331.251, "dur": 0.990,
    "args": {
      "External id": 2044,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1204, "finished": true, "Ev Idx": 2043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564833335.325, "dur": 1.137,
    "args": {
      "External id": 2045,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1205, "finished": true, "Ev Idx": 2044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564833337.962, "dur": 1.884,
    "args": {
      "External id": 2046,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1206, "finished": true, "Ev Idx": 2045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833338.808, "dur": 0.396,
    "args": {
      "External id": 2047,"Record function id": 0, "finished": true, "Ev Idx": 2046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564833343.253, "dur": 32.999,
    "args": {
      "External id": 2048,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1207, "finished": true, "Ev Idx": 2047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833375.338, "dur": 0.314,
    "args": {
      "External id": 2049,"Record function id": 0, "finished": true, "Ev Idx": 2048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564833377.481, "dur": 1.186,
    "args": {
      "External id": 2050,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1207, "finished": true, "Ev Idx": 2049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833377.922, "dur": 0.359,
    "args": {
      "External id": 2051,"Record function id": 0, "finished": true, "Ev Idx": 2050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564833379.534, "dur": 11.078,
    "args": {
      "External id": 2052,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1207, "finished": true, "Ev Idx": 2051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833396.992, "dur": 5.870,
    "args": {
      "External id": 2053,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1208, "finished": true, "Ev Idx": 2052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833400.270, "dur": 0.962,
    "args": {
      "External id": 2054,"Record function id": 0, "finished": true, "Ev Idx": 2053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833405.188, "dur": 1.822,
    "args": {
      "External id": 2055,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1209, "finished": true, "Ev Idx": 2054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833406.051, "dur": 0.327,
    "args": {
      "External id": 2056,"Record function id": 0, "finished": true, "Ev Idx": 2055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564833409.103, "dur": 12.752,
    "args": {
      "External id": 2057,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1210, "finished": true, "Ev Idx": 2056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564833426.820, "dur": 18.607,
    "args": {
      "External id": 2058,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1211, "finished": true, "Ev Idx": 2057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564833448.384, "dur": 8.718,
    "args": {
      "External id": 2059,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1212, "finished": true, "Ev Idx": 2058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564833459.179, "dur": 9.370,
    "args": {
      "External id": 2060,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1213, "finished": true, "Ev Idx": 2059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564833470.803, "dur": 7.330,
    "args": {
      "External id": 2061,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1214, "finished": true, "Ev Idx": 2060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833480.593, "dur": 2.163,
    "args": {
      "External id": 2062,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1215, "finished": true, "Ev Idx": 2061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833481.610, "dur": 0.461,
    "args": {
      "External id": 2063,"Record function id": 0, "finished": true, "Ev Idx": 2062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833484.323, "dur": 1.653,
    "args": {
      "External id": 2064,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1216, "finished": true, "Ev Idx": 2063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833485.127, "dur": 0.324,
    "args": {
      "External id": 2065,"Record function id": 0, "finished": true, "Ev Idx": 2064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564833487.033, "dur": 8.166,
    "args": {
      "External id": 2066,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1217, "finished": true, "Ev Idx": 2065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564833496.968, "dur": 9.838,
    "args": {
      "External id": 2067,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1218, "finished": true, "Ev Idx": 2066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564833509.019, "dur": 8.366,
    "args": {
      "External id": 2068,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1219, "finished": true, "Ev Idx": 2067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564833518.823, "dur": 8.346,
    "args": {
      "External id": 2069,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1220, "finished": true, "Ev Idx": 2068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833552.362, "dur": 2.573,
    "args": {
      "External id": 2070,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1221, "finished": true, "Ev Idx": 2069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833553.665, "dur": 0.469,
    "args": {
      "External id": 2071,"Record function id": 0, "finished": true, "Ev Idx": 2070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833555.588, "dur": 1.494,
    "args": {
      "External id": 2072,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1222, "finished": true, "Ev Idx": 2071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833556.281, "dur": 0.325,
    "args": {
      "External id": 2073,"Record function id": 0, "finished": true, "Ev Idx": 2072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564833557.805, "dur": 3.490,
    "args": {
      "External id": 2074,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1223, "finished": true, "Ev Idx": 2073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833559.704, "dur": 0.858,
    "args": {
      "External id": 2075,"Record function id": 0, "finished": true, "Ev Idx": 2074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833561.992, "dur": 2.495,
    "args": {
      "External id": 2076,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1224, "finished": true, "Ev Idx": 2075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833563.620, "dur": 0.352,
    "args": {
      "External id": 2077,"Record function id": 0, "finished": true, "Ev Idx": 2076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833565.097, "dur": 2.269,
    "args": {
      "External id": 2078,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1225, "finished": true, "Ev Idx": 2077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833566.569, "dur": 0.316,
    "args": {
      "External id": 2079,"Record function id": 0, "finished": true, "Ev Idx": 2078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564833570.542, "dur": 3.544,
    "args": {
      "External id": 2080,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1226, "finished": true, "Ev Idx": 2079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833572.965, "dur": 0.397,
    "args": {
      "External id": 2081,"Record function id": 0, "finished": true, "Ev Idx": 2080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833576.737, "dur": 21.536,
    "args": {
      "External id": 2082,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1227, "finished": true, "Ev Idx": 2081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564833579.667, "dur": 16.510,
    "args": {
      "External id": 2083,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1227, "finished": true, "Ev Idx": 2082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564833581.447, "dur": 3.007,
    "args": {
      "External id": 2084,"Record function id": 0, "finished": true, "Ev Idx": 2083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564833582.451, "dur": 1.823,
    "args": {
      "External id": 2085,"Record function id": 0, "finished": true, "Ev Idx": 2084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564833585.144, "dur": 10.378,
    "args": {
      "External id": 2086,"Record function id": 0, "finished": true, "Ev Idx": 2085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564833596.709, "dur": 1.301,
    "args": {
      "External id": 2087,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1228, "finished": true, "Ev Idx": 2086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833601.662, "dur": 2.592,
    "args": {
      "External id": 2088,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1229, "finished": true, "Ev Idx": 2087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833603.247, "dur": 0.421,
    "args": {
      "External id": 2089,"Record function id": 0, "finished": true, "Ev Idx": 2088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833604.848, "dur": 2.963,
    "args": {
      "External id": 2090,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1230, "finished": true, "Ev Idx": 2089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833606.851, "dur": 0.433,
    "args": {
      "External id": 2091,"Record function id": 0, "finished": true, "Ev Idx": 2090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564833608.382, "dur": 2.068,
    "args": {
      "External id": 2092,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1231, "finished": true, "Ev Idx": 2091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833609.658, "dur": 0.321,
    "args": {
      "External id": 2093,"Record function id": 0, "finished": true, "Ev Idx": 2092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833611.035, "dur": 1.513,
    "args": {
      "External id": 2094,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1232, "finished": true, "Ev Idx": 2093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833611.745, "dur": 0.316,
    "args": {
      "External id": 2095,"Record function id": 0, "finished": true, "Ev Idx": 2094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833613.148, "dur": 1.436,
    "args": {
      "External id": 2096,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1233, "finished": true, "Ev Idx": 2095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833613.838, "dur": 0.276,
    "args": {
      "External id": 2097,"Record function id": 0, "finished": true, "Ev Idx": 2096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564833616.136, "dur": 1.648,
    "args": {
      "External id": 2098,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1234, "finished": true, "Ev Idx": 2097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833617.017, "dur": 0.341,
    "args": {
      "External id": 2099,"Record function id": 0, "finished": true, "Ev Idx": 2098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833619.184, "dur": 12.814,
    "args": {
      "External id": 2100,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1235, "finished": true, "Ev Idx": 2099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564833620.103, "dur": 10.096,
    "args": {
      "External id": 2101,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1235, "finished": true, "Ev Idx": 2100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564833620.806, "dur": 1.675,
    "args": {
      "External id": 2102,"Record function id": 0, "finished": true, "Ev Idx": 2101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564833621.155, "dur": 1.212,
    "args": {
      "External id": 2103,"Record function id": 0, "finished": true, "Ev Idx": 2102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564833622.788, "dur": 7.058,
    "args": {
      "External id": 2104,"Record function id": 0, "finished": true, "Ev Idx": 2103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564833630.576, "dur": 1.261,
    "args": {
      "External id": 2105,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1236, "finished": true, "Ev Idx": 2104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564833634.205, "dur": 1.711,
    "args": {
      "External id": 2106,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1237, "finished": true, "Ev Idx": 2105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833635.041, "dur": 0.408,
    "args": {
      "External id": 2107,"Record function id": 0, "finished": true, "Ev Idx": 2106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564833637.547, "dur": 47.120,
    "args": {
      "External id": 2108,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1238, "finished": true, "Ev Idx": 2107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564833638.248, "dur": 1.658,
    "args": {
      "External id": 2109,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1238, "finished": true, "Ev Idx": 2108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833639.119, "dur": 0.338,
    "args": {
      "External id": 2110,"Record function id": 0, "finished": true, "Ev Idx": 2109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833640.294, "dur": 4.101,
    "args": {
      "External id": 2111,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1239, "finished": true, "Ev Idx": 2110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564833642.148, "dur": 2.099,
    "args": {
      "External id": 2112,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1239, "finished": true, "Ev Idx": 2111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564833644.907, "dur": 1.378,
    "args": {
      "External id": 2113,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1240, "finished": true, "Ev Idx": 2112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833645.610, "dur": 0.284,
    "args": {
      "External id": 2114,"Record function id": 0, "finished": true, "Ev Idx": 2113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833646.575, "dur": 2.661,
    "args": {
      "External id": 2115,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1241, "finished": true, "Ev Idx": 2114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564833647.420, "dur": 1.715,
    "args": {
      "External id": 2116,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1241, "finished": true, "Ev Idx": 2115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564833649.898, "dur": 32.646,
    "args": {
      "External id": 2117,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1242, "finished": true, "Ev Idx": 2116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564833664.283, "dur": 0.948,
    "args": {
      "External id": 2118,"Record function id": 0, "finished": true, "Ev Idx": 2117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564833683.208, "dur": 1.087,
    "args": {
      "External id": 2119,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1243, "finished": true, "Ev Idx": 2118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564833688.367, "dur": 10.070,
    "args": {
      "External id": 2120,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1244, "finished": true, "Ev Idx": 2119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833701.426, "dur": 1.976,
    "args": {
      "External id": 2121,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1245, "finished": true, "Ev Idx": 2120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833702.244, "dur": 0.502,
    "args": {
      "External id": 2122,"Record function id": 0, "finished": true, "Ev Idx": 2121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833704.183, "dur": 1.325,
    "args": {
      "External id": 2123,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1245, "finished": true, "Ev Idx": 2122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833704.749, "dur": 0.285,
    "args": {
      "External id": 2124,"Record function id": 0, "finished": true, "Ev Idx": 2123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564833706.136, "dur": 0.972,
    "args": {
      "External id": 2125,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1245, "finished": true, "Ev Idx": 2124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833706.549, "dur": 0.250,
    "args": {
      "External id": 2126,"Record function id": 0, "finished": true, "Ev Idx": 2125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564833708.651, "dur": 7.373,
    "args": {
      "External id": 2127,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1245, "finished": true, "Ev Idx": 2126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564833721.906, "dur": 14.898,
    "args": {
      "External id": 2128,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1246, "finished": true, "Ev Idx": 2127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564833722.794, "dur": 0.404,
    "args": {
      "External id": 2129,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1246, "finished": true, "Ev Idx": 2128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564833723.813, "dur": 12.716,
    "args": {
      "External id": 2130,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1246, "finished": true, "Ev Idx": 2129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564833739.870, "dur": 0.193,
    "args": {
      "External id": 2131,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1247, "finished": true, "Ev Idx": 2130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564833751.630, "dur": 0.505,
    "args": {
      "External id": 2132,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1247, "finished": true, "Ev Idx": 2131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564833753.788, "dur": 39.407,
    "args": {
      "External id": 2133,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1247, "finished": true, "Ev Idx": 2132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564833754.359, "dur": 2.661,
    "args": {
      "External id": 2134,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1247, "finished": true, "Ev Idx": 2133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833756.182, "dur": 0.348,
    "args": {
      "External id": 2135,"Record function id": 0, "finished": true, "Ev Idx": 2134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833757.413, "dur": 2.561,
    "args": {
      "External id": 2136,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1248, "finished": true, "Ev Idx": 2135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564833757.874, "dur": 1.972,
    "args": {
      "External id": 2137,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1248, "finished": true, "Ev Idx": 2136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564833760.535, "dur": 1.227,
    "args": {
      "External id": 2138,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1249, "finished": true, "Ev Idx": 2137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833761.137, "dur": 0.264,
    "args": {
      "External id": 2139,"Record function id": 0, "finished": true, "Ev Idx": 2138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833762.048, "dur": 1.158,
    "args": {
      "External id": 2140,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1250, "finished": true, "Ev Idx": 2139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564833762.298, "dur": 0.833,
    "args": {
      "External id": 2141,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1250, "finished": true, "Ev Idx": 2140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564833763.649, "dur": 27.535,
    "args": {
      "External id": 2142,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1251, "finished": true, "Ev Idx": 2141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564833774.037, "dur": 1.080,
    "args": {
      "External id": 2143,"Record function id": 0, "finished": true, "Ev Idx": 2142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564833791.708, "dur": 1.150,
    "args": {
      "External id": 2144,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1252, "finished": true, "Ev Idx": 2143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564833795.327, "dur": 1.906,
    "args": {
      "External id": 2145,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1253, "finished": true, "Ev Idx": 2144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833796.318, "dur": 0.393,
    "args": {
      "External id": 2146,"Record function id": 0, "finished": true, "Ev Idx": 2145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564833800.027, "dur": 11.558,
    "args": {
      "External id": 2147,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1254, "finished": true, "Ev Idx": 2146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564833800.492, "dur": 10.940,
    "args": {
      "External id": 2148,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1254, "finished": true, "Ev Idx": 2147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564833801.590, "dur": 2.417,
    "args": {
      "External id": 2149,"Record function id": 0, "finished": true, "Ev Idx": 2148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564833801.977, "dur": 1.883,
    "args": {
      "External id": 2150,"Record function id": 0, "finished": true, "Ev Idx": 2149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564833804.293, "dur": 6.630,
    "args": {
      "External id": 2151,"Record function id": 0, "finished": true, "Ev Idx": 2150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833817.288, "dur": 1.724,
    "args": {
      "External id": 2152,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1255, "finished": true, "Ev Idx": 2151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564833817.619, "dur": 1.299,
    "args": {
      "External id": 2153,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1255, "finished": true, "Ev Idx": 2152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564833832.264, "dur": 35.952,
    "args": {
      "External id": 2154,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1256, "finished": true, "Ev Idx": 2153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564833832.840, "dur": 2.802,
    "args": {
      "External id": 2155,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1256, "finished": true, "Ev Idx": 2154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564833834.197, "dur": 0.897,
    "args": {
      "External id": 2156,"Record function id": 0, "finished": true, "Ev Idx": 2155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564833834.658, "dur": 0.336,
    "args": {
      "External id": 2157,"Record function id": 0, "finished": true, "Ev Idx": 2156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564833835.987, "dur": 32.086,
    "args": {
      "External id": 2158,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1257, "finished": true, "Ev Idx": 2157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564833836.437, "dur": 1.435,
    "args": {
      "External id": 2159,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1257, "finished": true, "Ev Idx": 2158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564833836.692, "dur": 1.071,
    "args": {
      "External id": 2160,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1257, "finished": true, "Ev Idx": 2159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564833838.251, "dur": 27.859,
    "args": {
      "External id": 2161,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1258, "finished": true, "Ev Idx": 2160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564833851.157, "dur": 1.155,
    "args": {
      "External id": 2162,"Record function id": 0, "finished": true, "Ev Idx": 2161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564833866.654, "dur": 1.076,
    "args": {
      "External id": 2163,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1259, "finished": true, "Ev Idx": 2162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564833873.466, "dur": 76.322,
    "args": {
      "External id": 2164,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1260, "finished": true, "Ev Idx": 2163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564833975.130, "dur": 0.305,
    "args": {
      "External id": 2165,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1261, "finished": true, "Ev Idx": 2164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564833978.000, "dur": 14.267,
    "args": {
      "External id": 2166,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1261, "finished": true, "Ev Idx": 2165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564833980.336, "dur": 0.546,
    "args": {
      "External id": 2167,"Record function id": 0, "finished": true, "Ev Idx": 2166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564833981.340, "dur": 0.201,
    "args": {
      "External id": 2168,"Record function id": 0, "finished": true, "Ev Idx": 2167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564833995.330, "dur": 14.859,
    "args": {
      "External id": 2169,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1262, "finished": true, "Ev Idx": 2168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564834014.120, "dur": 8.721,
    "args": {
      "External id": 2170,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1263, "finished": true, "Ev Idx": 2169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564834024.999, "dur": 10.470,
    "args": {
      "External id": 2171,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1264, "finished": true, "Ev Idx": 2170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564834037.397, "dur": 8.763,
    "args": {
      "External id": 2172,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1265, "finished": true, "Ev Idx": 2171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564834048.747, "dur": 0.301,
    "args": {
      "External id": 2173,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1266, "finished": true, "Ev Idx": 2172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564834049.880, "dur": 7.270,
    "args": {
      "External id": 2174,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1266, "finished": true, "Ev Idx": 2173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564834096.593, "dur": 63.921,
    "args": {
      "External id": 2175,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1267, "finished": true, "Ev Idx": 2174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564834097.549, "dur": 5.996,
    "args": {
      "External id": 2176,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1267, "finished": true, "Ev Idx": 2175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564834099.879, "dur": 2.535,
    "args": {
      "External id": 2177,"Record function id": 0, "finished": true, "Ev Idx": 2176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834101.113, "dur": 0.884,
    "args": {
      "External id": 2178,"Record function id": 0, "finished": true, "Ev Idx": 2177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564834104.109, "dur": 56.237,
    "args": {
      "External id": 2179,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1268, "finished": true, "Ev Idx": 2178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564834104.906, "dur": 2.407,
    "args": {
      "External id": 2180,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1268, "finished": true, "Ev Idx": 2179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564834105.280, "dur": 1.903,
    "args": {
      "External id": 2181,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1268, "finished": true, "Ev Idx": 2180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564834107.747, "dur": 49.823,
    "args": {
      "External id": 2182,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1269, "finished": true, "Ev Idx": 2181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564834121.746, "dur": 1.463,
    "args": {
      "External id": 2183,"Record function id": 0, "finished": true, "Ev Idx": 2182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564834158.401, "dur": 1.492,
    "args": {
      "External id": 2184,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1270, "finished": true, "Ev Idx": 2183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564834170.713, "dur": 11.574,
    "args": {
      "External id": 2185,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1271, "finished": true, "Ev Idx": 2184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564834190.381, "dur": 95.746,
    "args": {
      "External id": 2186,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1272, "finished": true, "Ev Idx": 2185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564834190.691, "dur": 2.929,
    "args": {
      "External id": 2187,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1272, "finished": true, "Ev Idx": 2186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564834191.908, "dur": 1.156,
    "args": {
      "External id": 2188,"Record function id": 0, "finished": true, "Ev Idx": 2187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834192.384, "dur": 0.537,
    "args": {
      "External id": 2189,"Record function id": 0, "finished": true, "Ev Idx": 2188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564834194.094, "dur": 91.883,
    "args": {
      "External id": 2190,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1273, "finished": true, "Ev Idx": 2189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564834194.899, "dur": 2.322,
    "args": {
      "External id": 2191,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1273, "finished": true, "Ev Idx": 2190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564834195.649, "dur": 1.450,
    "args": {
      "External id": 2192,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1273, "finished": true, "Ev Idx": 2191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564834197.692, "dur": 85.991,
    "args": {
      "External id": 2193,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1274, "finished": true, "Ev Idx": 2192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564834257.377, "dur": 1.221,
    "args": {
      "External id": 2194,"Record function id": 0, "finished": true, "Ev Idx": 2193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564834284.422, "dur": 1.207,
    "args": {
      "External id": 2195,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1275, "finished": true, "Ev Idx": 2194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564834288.947, "dur": 9.628,
    "args": {
      "External id": 2196,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1276, "finished": true, "Ev Idx": 2195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564834306.361, "dur": 43.922,
    "args": {
      "External id": 2197,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1277, "finished": true, "Ev Idx": 2196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564834306.864, "dur": 3.176,
    "args": {
      "External id": 2198,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1277, "finished": true, "Ev Idx": 2197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564834307.981, "dur": 1.451,
    "args": {
      "External id": 2199,"Record function id": 0, "finished": true, "Ev Idx": 2198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834308.611, "dur": 0.694,
    "args": {
      "External id": 2200,"Record function id": 0, "finished": true, "Ev Idx": 2199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564834310.357, "dur": 39.824,
    "args": {
      "External id": 2201,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1278, "finished": true, "Ev Idx": 2200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564834310.920, "dur": 1.609,
    "args": {
      "External id": 2202,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1278, "finished": true, "Ev Idx": 2201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564834311.280, "dur": 1.159,
    "args": {
      "External id": 2203,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1278, "finished": true, "Ev Idx": 2202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564834312.851, "dur": 35.288,
    "args": {
      "External id": 2204,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1279, "finished": true, "Ev Idx": 2203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564834325.735, "dur": 1.132,
    "args": {
      "External id": 2205,"Record function id": 0, "finished": true, "Ev Idx": 2204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564834348.705, "dur": 1.157,
    "args": {
      "External id": 2206,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1280, "finished": true, "Ev Idx": 2205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564834353.466, "dur": 8.589,
    "args": {
      "External id": 2207,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1281, "finished": true, "Ev Idx": 2206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564834383.550, "dur": 0.252,
    "args": {
      "External id": 2208,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1282, "finished": true, "Ev Idx": 2207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564834385.981, "dur": 10.377,
    "args": {
      "External id": 2209,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1282, "finished": true, "Ev Idx": 2208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564834387.600, "dur": 0.504,
    "args": {
      "External id": 2210,"Record function id": 0, "finished": true, "Ev Idx": 2209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564834388.460, "dur": 0.068,
    "args": {
      "External id": 2211,"Record function id": 0, "finished": true, "Ev Idx": 2210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564834399.062, "dur": 12.747,
    "args": {
      "External id": 2212,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1283, "finished": true, "Ev Idx": 2211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564834416.020, "dur": 8.635,
    "args": {
      "External id": 2213,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1284, "finished": true, "Ev Idx": 2212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564834426.463, "dur": 10.184,
    "args": {
      "External id": 2214,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1285, "finished": true, "Ev Idx": 2213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564834438.311, "dur": 9.026,
    "args": {
      "External id": 2215,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1286, "finished": true, "Ev Idx": 2214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564834449.911, "dur": 0.242,
    "args": {
      "External id": 2216,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1287, "finished": true, "Ev Idx": 2215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564834451.142, "dur": 45.370,
    "args": {
      "External id": 2217,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1287, "finished": true, "Ev Idx": 2216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564834517.509, "dur": 47.330,
    "args": {
      "External id": 2218,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1288, "finished": true, "Ev Idx": 2217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564834518.445, "dur": 4.510,
    "args": {
      "External id": 2219,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1288, "finished": true, "Ev Idx": 2218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564834520.266, "dur": 1.648,
    "args": {
      "External id": 2220,"Record function id": 0, "finished": true, "Ev Idx": 2219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834521.072, "dur": 0.647,
    "args": {
      "External id": 2221,"Record function id": 0, "finished": true, "Ev Idx": 2220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564834523.354, "dur": 41.377,
    "args": {
      "External id": 2222,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1289, "finished": true, "Ev Idx": 2221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564834523.961, "dur": 3.110,
    "args": {
      "External id": 2223,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1289, "finished": true, "Ev Idx": 2222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564834524.332, "dur": 2.617,
    "args": {
      "External id": 2224,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1289, "finished": true, "Ev Idx": 2223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564834527.476, "dur": 34.590,
    "args": {
      "External id": 2225,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1290, "finished": true, "Ev Idx": 2224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564834540.772, "dur": 1.455,
    "args": {
      "External id": 2226,"Record function id": 0, "finished": true, "Ev Idx": 2225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564834562.861, "dur": 1.452,
    "args": {
      "External id": 2227,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1291, "finished": true, "Ev Idx": 2226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564834568.923, "dur": 1.462,
    "args": {
      "External id": 2228,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1292, "finished": true, "Ev Idx": 2227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564834572.342, "dur": 3.212,
    "args": {
      "External id": 2229,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1293, "finished": true, "Ev Idx": 2228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834573.764, "dur": 0.823,
    "args": {
      "External id": 2230,"Record function id": 0, "finished": true, "Ev Idx": 2229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564834582.744, "dur": 44.449,
    "args": {
      "External id": 2231,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1294, "finished": true, "Ev Idx": 2230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564834583.062, "dur": 2.468,
    "args": {
      "External id": 2232,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1294, "finished": true, "Ev Idx": 2231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564834584.064, "dur": 0.987,
    "args": {
      "External id": 2233,"Record function id": 0, "finished": true, "Ev Idx": 2232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834584.574, "dur": 0.381,
    "args": {
      "External id": 2234,"Record function id": 0, "finished": true, "Ev Idx": 2233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564834585.884, "dur": 41.211,
    "args": {
      "External id": 2235,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1295, "finished": true, "Ev Idx": 2234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564834586.565, "dur": 1.533,
    "args": {
      "External id": 2236,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1295, "finished": true, "Ev Idx": 2235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564834586.845, "dur": 1.155,
    "args": {
      "External id": 2237,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1295, "finished": true, "Ev Idx": 2236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564834588.409, "dur": 36.672,
    "args": {
      "External id": 2238,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1296, "finished": true, "Ev Idx": 2237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564834599.929, "dur": 1.131,
    "args": {
      "External id": 2239,"Record function id": 0, "finished": true, "Ev Idx": 2238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564834625.670, "dur": 1.149,
    "args": {
      "External id": 2240,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1297, "finished": true, "Ev Idx": 2239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564834629.600, "dur": 1.194,
    "args": {
      "External id": 2241,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1298, "finished": true, "Ev Idx": 2240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564834632.490, "dur": 1.890,
    "args": {
      "External id": 2242,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1299, "finished": true, "Ev Idx": 2241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834633.234, "dur": 0.492,
    "args": {
      "External id": 2243,"Record function id": 0, "finished": true, "Ev Idx": 2242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564834641.361, "dur": 38.075,
    "args": {
      "External id": 2244,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1300, "finished": true, "Ev Idx": 2243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564834641.673, "dur": 2.136,
    "args": {
      "External id": 2245,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1300, "finished": true, "Ev Idx": 2244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564834642.429, "dur": 0.917,
    "args": {
      "External id": 2246,"Record function id": 0, "finished": true, "Ev Idx": 2245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834642.856, "dur": 0.390,
    "args": {
      "External id": 2247,"Record function id": 0, "finished": true, "Ev Idx": 2246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564834644.100, "dur": 35.239,
    "args": {
      "External id": 2248,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1301, "finished": true, "Ev Idx": 2247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564834644.646, "dur": 1.392,
    "args": {
      "External id": 2249,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1301, "finished": true, "Ev Idx": 2248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564834645.115, "dur": 0.826,
    "args": {
      "External id": 2250,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1301, "finished": true, "Ev Idx": 2249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564834646.466, "dur": 31.127,
    "args": {
      "External id": 2251,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1302, "finished": true, "Ev Idx": 2250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564834655.898, "dur": 1.114,
    "args": {
      "External id": 2252,"Record function id": 0, "finished": true, "Ev Idx": 2251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564834678.098, "dur": 0.970,
    "args": {
      "External id": 2253,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1303, "finished": true, "Ev Idx": 2252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564834681.727, "dur": 1.145,
    "args": {
      "External id": 2254,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1304, "finished": true, "Ev Idx": 2253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564834684.284, "dur": 1.559,
    "args": {
      "External id": 2255,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1305, "finished": true, "Ev Idx": 2254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834684.985, "dur": 0.352,
    "args": {
      "External id": 2256,"Record function id": 0, "finished": true, "Ev Idx": 2255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564834688.959, "dur": 2.338,
    "args": {
      "External id": 2257,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1306, "finished": true, "Ev Idx": 2256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834690.440, "dur": 0.318,
    "args": {
      "External id": 2258,"Record function id": 0, "finished": true, "Ev Idx": 2257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564834692.478, "dur": 1.080,
    "args": {
      "External id": 2259,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1306, "finished": true, "Ev Idx": 2258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834692.892, "dur": 0.302,
    "args": {
      "External id": 2260,"Record function id": 0, "finished": true, "Ev Idx": 2259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564834694.384, "dur": 10.595,
    "args": {
      "External id": 2261,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1306, "finished": true, "Ev Idx": 2260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834711.587, "dur": 5.029,
    "args": {
      "External id": 2262,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1307, "finished": true, "Ev Idx": 2261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834714.357, "dur": 0.554,
    "args": {
      "External id": 2263,"Record function id": 0, "finished": true, "Ev Idx": 2262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834719.024, "dur": 2.597,
    "args": {
      "External id": 2264,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1308, "finished": true, "Ev Idx": 2263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834720.664, "dur": 0.347,
    "args": {
      "External id": 2265,"Record function id": 0, "finished": true, "Ev Idx": 2264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564834723.535, "dur": 10.182,
    "args": {
      "External id": 2266,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1309, "finished": true, "Ev Idx": 2265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564834738.438, "dur": 18.388,
    "args": {
      "External id": 2267,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1310, "finished": true, "Ev Idx": 2266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564834760.123, "dur": 8.461,
    "args": {
      "External id": 2268,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1311, "finished": true, "Ev Idx": 2267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564834770.271, "dur": 9.551,
    "args": {
      "External id": 2269,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1312, "finished": true, "Ev Idx": 2268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564834781.827, "dur": 9.309,
    "args": {
      "External id": 2270,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1313, "finished": true, "Ev Idx": 2269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834793.574, "dur": 2.255,
    "args": {
      "External id": 2271,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1314, "finished": true, "Ev Idx": 2270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834794.735, "dur": 0.446,
    "args": {
      "External id": 2272,"Record function id": 0, "finished": true, "Ev Idx": 2271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834797.361, "dur": 2.113,
    "args": {
      "External id": 2273,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1315, "finished": true, "Ev Idx": 2272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834798.489, "dur": 0.471,
    "args": {
      "External id": 2274,"Record function id": 0, "finished": true, "Ev Idx": 2273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564834800.509, "dur": 8.157,
    "args": {
      "External id": 2275,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1316, "finished": true, "Ev Idx": 2274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564834810.581, "dur": 9.315,
    "args": {
      "External id": 2276,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1317, "finished": true, "Ev Idx": 2275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564834822.279, "dur": 7.991,
    "args": {
      "External id": 2277,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1318, "finished": true, "Ev Idx": 2276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564834831.652, "dur": 6.389,
    "args": {
      "External id": 2278,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1319, "finished": true, "Ev Idx": 2277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834863.539, "dur": 2.655,
    "args": {
      "External id": 2279,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1320, "finished": true, "Ev Idx": 2278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834864.957, "dur": 0.526,
    "args": {
      "External id": 2280,"Record function id": 0, "finished": true, "Ev Idx": 2279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834866.992, "dur": 1.654,
    "args": {
      "External id": 2281,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1321, "finished": true, "Ev Idx": 2280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834867.703, "dur": 0.452,
    "args": {
      "External id": 2282,"Record function id": 0, "finished": true, "Ev Idx": 2281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564834869.517, "dur": 2.432,
    "args": {
      "External id": 2283,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1322, "finished": true, "Ev Idx": 2282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834870.364, "dur": 0.912,
    "args": {
      "External id": 2284,"Record function id": 0, "finished": true, "Ev Idx": 2283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834872.554, "dur": 3.132,
    "args": {
      "External id": 2285,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1323, "finished": true, "Ev Idx": 2284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834874.778, "dur": 0.369,
    "args": {
      "External id": 2286,"Record function id": 0, "finished": true, "Ev Idx": 2285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834876.256, "dur": 2.262,
    "args": {
      "External id": 2287,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1324, "finished": true, "Ev Idx": 2286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834877.605, "dur": 0.432,
    "args": {
      "External id": 2288,"Record function id": 0, "finished": true, "Ev Idx": 2287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564834881.467, "dur": 3.305,
    "args": {
      "External id": 2289,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1325, "finished": true, "Ev Idx": 2288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834883.599, "dur": 0.412,
    "args": {
      "External id": 2290,"Record function id": 0, "finished": true, "Ev Idx": 2289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564834887.413, "dur": 21.015,
    "args": {
      "External id": 2291,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1326, "finished": true, "Ev Idx": 2290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564834890.146, "dur": 16.204,
    "args": {
      "External id": 2292,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1326, "finished": true, "Ev Idx": 2291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564834891.933, "dur": 3.132,
    "args": {
      "External id": 2293,"Record function id": 0, "finished": true, "Ev Idx": 2292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564834892.971, "dur": 1.837,
    "args": {
      "External id": 2294,"Record function id": 0, "finished": true, "Ev Idx": 2293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564834895.759, "dur": 9.970,
    "args": {
      "External id": 2295,"Record function id": 0, "finished": true, "Ev Idx": 2294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564834906.890, "dur": 1.263,
    "args": {
      "External id": 2296,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1327, "finished": true, "Ev Idx": 2295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834911.834, "dur": 2.882,
    "args": {
      "External id": 2297,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1328, "finished": true, "Ev Idx": 2296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834913.728, "dur": 0.424,
    "args": {
      "External id": 2298,"Record function id": 0, "finished": true, "Ev Idx": 2297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834915.310, "dur": 2.259,
    "args": {
      "External id": 2299,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1329, "finished": true, "Ev Idx": 2298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834916.793, "dur": 0.287,
    "args": {
      "External id": 2300,"Record function id": 0, "finished": true, "Ev Idx": 2299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564834918.103, "dur": 1.429,
    "args": {
      "External id": 2301,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1330, "finished": true, "Ev Idx": 2300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834918.798, "dur": 0.313,
    "args": {
      "External id": 2302,"Record function id": 0, "finished": true, "Ev Idx": 2301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834920.131, "dur": 1.735,
    "args": {
      "External id": 2303,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1331, "finished": true, "Ev Idx": 2302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834920.995, "dur": 0.391,
    "args": {
      "External id": 2304,"Record function id": 0, "finished": true, "Ev Idx": 2303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564834922.569, "dur": 1.393,
    "args": {
      "External id": 2305,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1332, "finished": true, "Ev Idx": 2304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834923.244, "dur": 0.267,
    "args": {
      "External id": 2306,"Record function id": 0, "finished": true, "Ev Idx": 2305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564834925.496, "dur": 1.693,
    "args": {
      "External id": 2307,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1333, "finished": true, "Ev Idx": 2306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834926.430, "dur": 0.317,
    "args": {
      "External id": 2308,"Record function id": 0, "finished": true, "Ev Idx": 2307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564834928.706, "dur": 13.690,
    "args": {
      "External id": 2309,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1334, "finished": true, "Ev Idx": 2308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564834929.815, "dur": 10.630,
    "args": {
      "External id": 2310,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1334, "finished": true, "Ev Idx": 2309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564834930.552, "dur": 1.590,
    "args": {
      "External id": 2311,"Record function id": 0, "finished": true, "Ev Idx": 2310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564834930.880, "dur": 1.156,
    "args": {
      "External id": 2312,"Record function id": 0, "finished": true, "Ev Idx": 2311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564834932.432, "dur": 7.682,
    "args": {
      "External id": 2313,"Record function id": 0, "finished": true, "Ev Idx": 2312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564834940.871, "dur": 1.359,
    "args": {
      "External id": 2314,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1335, "finished": true, "Ev Idx": 2313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564834946.011, "dur": 1.803,
    "args": {
      "External id": 2315,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1336, "finished": true, "Ev Idx": 2314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834946.881, "dur": 0.399,
    "args": {
      "External id": 2316,"Record function id": 0, "finished": true, "Ev Idx": 2315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564834949.553, "dur": 49.087,
    "args": {
      "External id": 2317,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1337, "finished": true, "Ev Idx": 2316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564834950.255, "dur": 1.662,
    "args": {
      "External id": 2318,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1337, "finished": true, "Ev Idx": 2317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834951.118, "dur": 0.344,
    "args": {
      "External id": 2319,"Record function id": 0, "finished": true, "Ev Idx": 2318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564834952.315, "dur": 4.316,
    "args": {
      "External id": 2320,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1338, "finished": true, "Ev Idx": 2319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564834954.514, "dur": 1.961,
    "args": {
      "External id": 2321,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1338, "finished": true, "Ev Idx": 2320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564834957.150, "dur": 1.609,
    "args": {
      "External id": 2322,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1339, "finished": true, "Ev Idx": 2321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564834958.080, "dur": 0.295,
    "args": {
      "External id": 2323,"Record function id": 0, "finished": true, "Ev Idx": 2322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564834959.067, "dur": 2.652,
    "args": {
      "External id": 2324,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1340, "finished": true, "Ev Idx": 2323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564834959.875, "dur": 1.751,
    "args": {
      "External id": 2325,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1340, "finished": true, "Ev Idx": 2324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564834962.447, "dur": 34.086,
    "args": {
      "External id": 2326,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1341, "finished": true, "Ev Idx": 2325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564834977.986, "dur": 1.164,
    "args": {
      "External id": 2327,"Record function id": 0, "finished": true, "Ev Idx": 2326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564834997.103, "dur": 1.097,
    "args": {
      "External id": 2328,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1342, "finished": true, "Ev Idx": 2327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564835002.292, "dur": 9.701,
    "args": {
      "External id": 2329,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1343, "finished": true, "Ev Idx": 2328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564835015.004, "dur": 1.989,
    "args": {
      "External id": 2330,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1344, "finished": true, "Ev Idx": 2329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835015.806, "dur": 0.543,
    "args": {
      "External id": 2331,"Record function id": 0, "finished": true, "Ev Idx": 2330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564835017.627, "dur": 1.307,
    "args": {
      "External id": 2332,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1344, "finished": true, "Ev Idx": 2331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835018.208, "dur": 0.288,
    "args": {
      "External id": 2333,"Record function id": 0, "finished": true, "Ev Idx": 2332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564835019.692, "dur": 1.196,
    "args": {
      "External id": 2334,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1344, "finished": true, "Ev Idx": 2333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835020.327, "dur": 0.254,
    "args": {
      "External id": 2335,"Record function id": 0, "finished": true, "Ev Idx": 2334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564835022.613, "dur": 7.335,
    "args": {
      "External id": 2336,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1344, "finished": true, "Ev Idx": 2335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564835035.821, "dur": 13.145,
    "args": {
      "External id": 2337,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1345, "finished": true, "Ev Idx": 2336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564835036.770, "dur": 0.371,
    "args": {
      "External id": 2338,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1345, "finished": true, "Ev Idx": 2337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564835037.885, "dur": 10.834,
    "args": {
      "External id": 2339,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1345, "finished": true, "Ev Idx": 2338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564835051.880, "dur": 0.392,
    "args": {
      "External id": 2340,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1346, "finished": true, "Ev Idx": 2339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564835063.871, "dur": 0.500,
    "args": {
      "External id": 2341,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1346, "finished": true, "Ev Idx": 2340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564835066.233, "dur": 38.451,
    "args": {
      "External id": 2342,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1346, "finished": true, "Ev Idx": 2341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564835066.765, "dur": 2.505,
    "args": {
      "External id": 2343,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1346, "finished": true, "Ev Idx": 2342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835068.372, "dur": 0.363,
    "args": {
      "External id": 2344,"Record function id": 0, "finished": true, "Ev Idx": 2343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564835069.781, "dur": 2.875,
    "args": {
      "External id": 2345,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1347, "finished": true, "Ev Idx": 2344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835070.243, "dur": 2.296,
    "args": {
      "External id": 2346,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1347, "finished": true, "Ev Idx": 2345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564835073.110, "dur": 1.318,
    "args": {
      "External id": 2347,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1348, "finished": true, "Ev Idx": 2346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835073.803, "dur": 0.271,
    "args": {
      "External id": 2348,"Record function id": 0, "finished": true, "Ev Idx": 2347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564835074.709, "dur": 1.347,
    "args": {
      "External id": 2349,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1349, "finished": true, "Ev Idx": 2348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835074.984, "dur": 0.998,
    "args": {
      "External id": 2350,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1349, "finished": true, "Ev Idx": 2349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564835076.458, "dur": 26.349,
    "args": {
      "External id": 2351,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1350, "finished": true, "Ev Idx": 2350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564835086.891, "dur": 1.044,
    "args": {
      "External id": 2352,"Record function id": 0, "finished": true, "Ev Idx": 2351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564835103.356, "dur": 1.046,
    "args": {
      "External id": 2353,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1351, "finished": true, "Ev Idx": 2352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835106.777, "dur": 2.175,
    "args": {
      "External id": 2354,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1352, "finished": true, "Ev Idx": 2353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835107.992, "dur": 0.391,
    "args": {
      "External id": 2355,"Record function id": 0, "finished": true, "Ev Idx": 2354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564835111.634, "dur": 11.877,
    "args": {
      "External id": 2356,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1353, "finished": true, "Ev Idx": 2355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564835112.241, "dur": 11.115,
    "args": {
      "External id": 2357,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1353, "finished": true, "Ev Idx": 2356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564835113.281, "dur": 2.020,
    "args": {
      "External id": 2358,"Record function id": 0, "finished": true, "Ev Idx": 2357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564835113.768, "dur": 1.389,
    "args": {
      "External id": 2359,"Record function id": 0, "finished": true, "Ev Idx": 2358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564835115.868, "dur": 7.075,
    "args": {
      "External id": 2360,"Record function id": 0, "finished": true, "Ev Idx": 2359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564835129.280, "dur": 1.901,
    "args": {
      "External id": 2361,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1354, "finished": true, "Ev Idx": 2360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835129.632, "dur": 1.463,
    "args": {
      "External id": 2362,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1354, "finished": true, "Ev Idx": 2361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564835144.005, "dur": 37.499,
    "args": {
      "External id": 2363,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1355, "finished": true, "Ev Idx": 2362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564835144.695, "dur": 3.370,
    "args": {
      "External id": 2364,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1355, "finished": true, "Ev Idx": 2363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835146.332, "dur": 1.150,
    "args": {
      "External id": 2365,"Record function id": 0, "finished": true, "Ev Idx": 2364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835147.032, "dur": 0.328,
    "args": {
      "External id": 2366,"Record function id": 0, "finished": true, "Ev Idx": 2365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564835148.594, "dur": 32.766,
    "args": {
      "External id": 2367,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1356, "finished": true, "Ev Idx": 2366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564835149.289, "dur": 1.348,
    "args": {
      "External id": 2368,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1356, "finished": true, "Ev Idx": 2367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835149.566, "dur": 0.976,
    "args": {
      "External id": 2369,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1356, "finished": true, "Ev Idx": 2368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564835151.031, "dur": 28.160,
    "args": {
      "External id": 2370,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1357, "finished": true, "Ev Idx": 2369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564835163.763, "dur": 1.129,
    "args": {
      "External id": 2371,"Record function id": 0, "finished": true, "Ev Idx": 2370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564835179.790, "dur": 1.223,
    "args": {
      "External id": 2372,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1358, "finished": true, "Ev Idx": 2371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564835186.467, "dur": 9.145,
    "args": {
      "External id": 2373,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1359, "finished": true, "Ev Idx": 2372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564835204.467, "dur": 0.195,
    "args": {
      "External id": 2374,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1360, "finished": true, "Ev Idx": 2373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564835206.721, "dur": 11.983,
    "args": {
      "External id": 2375,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1360, "finished": true, "Ev Idx": 2374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564835208.536, "dur": 0.369,
    "args": {
      "External id": 2376,"Record function id": 0, "finished": true, "Ev Idx": 2375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564835209.321, "dur": 0.070,
    "args": {
      "External id": 2377,"Record function id": 0, "finished": true, "Ev Idx": 2376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564835221.245, "dur": 11.714,
    "args": {
      "External id": 2378,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1361, "finished": true, "Ev Idx": 2377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564835236.779, "dur": 9.446,
    "args": {
      "External id": 2379,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1362, "finished": true, "Ev Idx": 2378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564835248.084, "dur": 10.674,
    "args": {
      "External id": 2380,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1363, "finished": true, "Ev Idx": 2379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564835260.522, "dur": 7.761,
    "args": {
      "External id": 2381,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1364, "finished": true, "Ev Idx": 2380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564835270.892, "dur": 0.266,
    "args": {
      "External id": 2382,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1365, "finished": true, "Ev Idx": 2381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564835272.234, "dur": 8.440,
    "args": {
      "External id": 2383,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1365, "finished": true, "Ev Idx": 2382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564835294.318, "dur": 108.564,
    "args": {
      "External id": 2384,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1366, "finished": true, "Ev Idx": 2383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564835295.003, "dur": 3.582,
    "args": {
      "External id": 2385,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1366, "finished": true, "Ev Idx": 2384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835296.375, "dur": 1.656,
    "args": {
      "External id": 2386,"Record function id": 0, "finished": true, "Ev Idx": 2385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835297.384, "dur": 0.526,
    "args": {
      "External id": 2387,"Record function id": 0, "finished": true, "Ev Idx": 2386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564835299.264, "dur": 103.383,
    "args": {
      "External id": 2388,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1367, "finished": true, "Ev Idx": 2387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564835299.972, "dur": 2.278,
    "args": {
      "External id": 2389,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1367, "finished": true, "Ev Idx": 2388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835300.564, "dur": 1.603,
    "args": {
      "External id": 2390,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1367, "finished": true, "Ev Idx": 2389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564835302.558, "dur": 97.828,
    "args": {
      "External id": 2391,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1368, "finished": true, "Ev Idx": 2390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564835370.596, "dur": 1.311,
    "args": {
      "External id": 2392,"Record function id": 0, "finished": true, "Ev Idx": 2391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564835401.051, "dur": 1.181,
    "args": {
      "External id": 2393,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1369, "finished": true, "Ev Idx": 2392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564835413.037, "dur": 10.399,
    "args": {
      "External id": 2394,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1370, "finished": true, "Ev Idx": 2393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564835431.953, "dur": 77.909,
    "args": {
      "External id": 2395,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1371, "finished": true, "Ev Idx": 2394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564835432.428, "dur": 3.365,
    "args": {
      "External id": 2396,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1371, "finished": true, "Ev Idx": 2395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835433.942, "dur": 1.211,
    "args": {
      "External id": 2397,"Record function id": 0, "finished": true, "Ev Idx": 2396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835434.646, "dur": 0.393,
    "args": {
      "External id": 2398,"Record function id": 0, "finished": true, "Ev Idx": 2397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564835436.229, "dur": 73.514,
    "args": {
      "External id": 2399,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1372, "finished": true, "Ev Idx": 2398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564835436.743, "dur": 1.681,
    "args": {
      "External id": 2400,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1372, "finished": true, "Ev Idx": 2399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835437.174, "dur": 1.138,
    "args": {
      "External id": 2401,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1372, "finished": true, "Ev Idx": 2400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564835438.713, "dur": 68.809,
    "args": {
      "External id": 2402,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1373, "finished": true, "Ev Idx": 2401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564835482.872, "dur": 1.175,
    "args": {
      "External id": 2403,"Record function id": 0, "finished": true, "Ev Idx": 2402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564835508.151, "dur": 1.272,
    "args": {
      "External id": 2404,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1374, "finished": true, "Ev Idx": 2403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564835512.080, "dur": 9.353,
    "args": {
      "External id": 2405,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1375, "finished": true, "Ev Idx": 2404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564835528.483, "dur": 38.124,
    "args": {
      "External id": 2406,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1376, "finished": true, "Ev Idx": 2405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564835528.968, "dur": 2.843,
    "args": {
      "External id": 2407,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1376, "finished": true, "Ev Idx": 2406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835530.176, "dur": 1.050,
    "args": {
      "External id": 2408,"Record function id": 0, "finished": true, "Ev Idx": 2407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835530.627, "dur": 0.405,
    "args": {
      "External id": 2409,"Record function id": 0, "finished": true, "Ev Idx": 2408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564835532.109, "dur": 34.392,
    "args": {
      "External id": 2410,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1377, "finished": true, "Ev Idx": 2409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564835532.516, "dur": 2.113,
    "args": {
      "External id": 2411,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1377, "finished": true, "Ev Idx": 2410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835532.799, "dur": 1.739,
    "args": {
      "External id": 2412,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1377, "finished": true, "Ev Idx": 2411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564835534.966, "dur": 29.665,
    "args": {
      "External id": 2413,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1378, "finished": true, "Ev Idx": 2412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564835543.548, "dur": 1.107,
    "args": {
      "External id": 2414,"Record function id": 0, "finished": true, "Ev Idx": 2413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564835565.328, "dur": 0.891,
    "args": {
      "External id": 2415,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1379, "finished": true, "Ev Idx": 2414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564835569.478, "dur": 9.460,
    "args": {
      "External id": 2416,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1380, "finished": true, "Ev Idx": 2415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564835598.567, "dur": 0.177,
    "args": {
      "External id": 2417,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1381, "finished": true, "Ev Idx": 2416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564835600.702, "dur": 10.842,
    "args": {
      "External id": 2418,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1381, "finished": true, "Ev Idx": 2417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564835602.225, "dur": 0.489,
    "args": {
      "External id": 2419,"Record function id": 0, "finished": true, "Ev Idx": 2418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564835603.190, "dur": 0.067,
    "args": {
      "External id": 2420,"Record function id": 0, "finished": true, "Ev Idx": 2419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564835614.071, "dur": 11.914,
    "args": {
      "External id": 2421,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1382, "finished": true, "Ev Idx": 2420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564835630.025, "dur": 7.533,
    "args": {
      "External id": 2422,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1383, "finished": true, "Ev Idx": 2421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564835639.223, "dur": 9.887,
    "args": {
      "External id": 2423,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1384, "finished": true, "Ev Idx": 2422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564835650.725, "dur": 8.278,
    "args": {
      "External id": 2424,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1385, "finished": true, "Ev Idx": 2423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564835661.535, "dur": 0.296,
    "args": {
      "External id": 2425,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1386, "finished": true, "Ev Idx": 2424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564835662.629, "dur": 9.333,
    "args": {
      "External id": 2426,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1386, "finished": true, "Ev Idx": 2425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564835688.961, "dur": 38.266,
    "args": {
      "External id": 2427,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1387, "finished": true, "Ev Idx": 2426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564835689.544, "dur": 3.741,
    "args": {
      "External id": 2428,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1387, "finished": true, "Ev Idx": 2427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835690.922, "dur": 1.443,
    "args": {
      "External id": 2429,"Record function id": 0, "finished": true, "Ev Idx": 2428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835691.813, "dur": 0.429,
    "args": {
      "External id": 2430,"Record function id": 0, "finished": true, "Ev Idx": 2429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564835693.718, "dur": 33.407,
    "args": {
      "External id": 2431,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1388, "finished": true, "Ev Idx": 2430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564835694.218, "dur": 2.478,
    "args": {
      "External id": 2432,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1388, "finished": true, "Ev Idx": 2431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835694.518, "dur": 2.058,
    "args": {
      "External id": 2433,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1388, "finished": true, "Ev Idx": 2432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564835697.005, "dur": 28.227,
    "args": {
      "External id": 2434,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1389, "finished": true, "Ev Idx": 2433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564835707.878, "dur": 1.082,
    "args": {
      "External id": 2435,"Record function id": 0, "finished": true, "Ev Idx": 2434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564835725.813, "dur": 0.998,
    "args": {
      "External id": 2436,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1390, "finished": true, "Ev Idx": 2435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835730.164, "dur": 1.324,
    "args": {
      "External id": 2437,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1391, "finished": true, "Ev Idx": 2436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835733.357, "dur": 2.329,
    "args": {
      "External id": 2438,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1392, "finished": true, "Ev Idx": 2437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835734.296, "dur": 0.557,
    "args": {
      "External id": 2439,"Record function id": 0, "finished": true, "Ev Idx": 2438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564835742.499, "dur": 39.165,
    "args": {
      "External id": 2440,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1393, "finished": true, "Ev Idx": 2439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564835742.839, "dur": 2.168,
    "args": {
      "External id": 2441,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1393, "finished": true, "Ev Idx": 2440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835743.724, "dur": 0.863,
    "args": {
      "External id": 2442,"Record function id": 0, "finished": true, "Ev Idx": 2441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835744.196, "dur": 0.289,
    "args": {
      "External id": 2443,"Record function id": 0, "finished": true, "Ev Idx": 2442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564835745.303, "dur": 36.262,
    "args": {
      "External id": 2444,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1394, "finished": true, "Ev Idx": 2443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564835745.876, "dur": 1.391,
    "args": {
      "External id": 2445,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1394, "finished": true, "Ev Idx": 2444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835746.297, "dur": 0.881,
    "args": {
      "External id": 2446,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1394, "finished": true, "Ev Idx": 2445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564835747.569, "dur": 32.088,
    "args": {
      "External id": 2447,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1395, "finished": true, "Ev Idx": 2446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564835756.834, "dur": 1.074,
    "args": {
      "External id": 2448,"Record function id": 0, "finished": true, "Ev Idx": 2447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564835780.242, "dur": 1.047,
    "args": {
      "External id": 2449,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1396, "finished": true, "Ev Idx": 2448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835783.801, "dur": 1.305,
    "args": {
      "External id": 2450,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1397, "finished": true, "Ev Idx": 2449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835786.518, "dur": 1.688,
    "args": {
      "External id": 2451,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1398, "finished": true, "Ev Idx": 2450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835787.262, "dur": 0.452,
    "args": {
      "External id": 2452,"Record function id": 0, "finished": true, "Ev Idx": 2451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564835795.114, "dur": 36.424,
    "args": {
      "External id": 2453,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1399, "finished": true, "Ev Idx": 2452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564835795.519, "dur": 2.080,
    "args": {
      "External id": 2454,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1399, "finished": true, "Ev Idx": 2453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835796.407, "dur": 0.786,
    "args": {
      "External id": 2455,"Record function id": 0, "finished": true, "Ev Idx": 2454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835796.824, "dur": 0.266,
    "args": {
      "External id": 2456,"Record function id": 0, "finished": true, "Ev Idx": 2455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564835797.915, "dur": 33.525,
    "args": {
      "External id": 2457,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1400, "finished": true, "Ev Idx": 2456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564835798.337, "dur": 1.457,
    "args": {
      "External id": 2458,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1400, "finished": true, "Ev Idx": 2457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835798.594, "dur": 1.130,
    "args": {
      "External id": 2459,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1400, "finished": true, "Ev Idx": 2458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564835800.110, "dur": 29.630,
    "args": {
      "External id": 2460,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1401, "finished": true, "Ev Idx": 2459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564835810.141, "dur": 1.088,
    "args": {
      "External id": 2461,"Record function id": 0, "finished": true, "Ev Idx": 2460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564835830.249, "dur": 0.930,
    "args": {
      "External id": 2462,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1402, "finished": true, "Ev Idx": 2461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564835833.931, "dur": 1.100,
    "args": {
      "External id": 2463,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1403, "finished": true, "Ev Idx": 2462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564835838.170, "dur": 1.753,
    "args": {
      "External id": 2464,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1404, "finished": true, "Ev Idx": 2463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835838.989, "dur": 0.325,
    "args": {
      "External id": 2465,"Record function id": 0, "finished": true, "Ev Idx": 2464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564835842.885, "dur": 2.176,
    "args": {
      "External id": 2466,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1405, "finished": true, "Ev Idx": 2465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835844.167, "dur": 0.313,
    "args": {
      "External id": 2467,"Record function id": 0, "finished": true, "Ev Idx": 2466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564835846.226, "dur": 1.128,
    "args": {
      "External id": 2468,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1405, "finished": true, "Ev Idx": 2467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835846.688, "dur": 0.306,
    "args": {
      "External id": 2469,"Record function id": 0, "finished": true, "Ev Idx": 2468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564835848.291, "dur": 11.016,
    "args": {
      "External id": 2470,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1405, "finished": true, "Ev Idx": 2469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564835864.542, "dur": 4.039,
    "args": {
      "External id": 2471,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1406, "finished": true, "Ev Idx": 2470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835866.692, "dur": 0.579,
    "args": {
      "External id": 2472,"Record function id": 0, "finished": true, "Ev Idx": 2471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564835870.914, "dur": 1.790,
    "args": {
      "External id": 2473,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1407, "finished": true, "Ev Idx": 2472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835871.764, "dur": 0.349,
    "args": {
      "External id": 2474,"Record function id": 0, "finished": true, "Ev Idx": 2473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564835874.411, "dur": 12.385,
    "args": {
      "External id": 2475,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1408, "finished": true, "Ev Idx": 2474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564835890.978, "dur": 57.156,
    "args": {
      "External id": 2476,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1409, "finished": true, "Ev Idx": 2475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564835951.612, "dur": 8.778,
    "args": {
      "External id": 2477,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1410, "finished": true, "Ev Idx": 2476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564835962.241, "dur": 8.187,
    "args": {
      "External id": 2478,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1411, "finished": true, "Ev Idx": 2477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564835972.786, "dur": 7.828,
    "args": {
      "External id": 2479,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1412, "finished": true, "Ev Idx": 2478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564835983.635, "dur": 3.222,
    "args": {
      "External id": 2480,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1413, "finished": true, "Ev Idx": 2479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835985.592, "dur": 0.431,
    "args": {
      "External id": 2481,"Record function id": 0, "finished": true, "Ev Idx": 2480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564835988.541, "dur": 1.697,
    "args": {
      "External id": 2482,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1414, "finished": true, "Ev Idx": 2481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564835989.371, "dur": 0.314,
    "args": {
      "External id": 2483,"Record function id": 0, "finished": true, "Ev Idx": 2482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564835991.479, "dur": 7.043,
    "args": {
      "External id": 2484,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1415, "finished": true, "Ev Idx": 2483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564836000.791, "dur": 12.466,
    "args": {
      "External id": 2485,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1416, "finished": true, "Ev Idx": 2484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564836015.507, "dur": 7.387,
    "args": {
      "External id": 2486,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1417, "finished": true, "Ev Idx": 2485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564836024.353, "dur": 6.680,
    "args": {
      "External id": 2487,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1418, "finished": true, "Ev Idx": 2486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836054.298, "dur": 3.445,
    "args": {
      "External id": 2488,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1419, "finished": true, "Ev Idx": 2487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836056.345, "dur": 0.630,
    "args": {
      "External id": 2489,"Record function id": 0, "finished": true, "Ev Idx": 2488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836058.411, "dur": 2.083,
    "args": {
      "External id": 2490,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1420, "finished": true, "Ev Idx": 2489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836059.611, "dur": 0.340,
    "args": {
      "External id": 2491,"Record function id": 0, "finished": true, "Ev Idx": 2490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564836061.180, "dur": 2.162,
    "args": {
      "External id": 2492,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1421, "finished": true, "Ev Idx": 2491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836062.124, "dur": 0.570,
    "args": {
      "External id": 2493,"Record function id": 0, "finished": true, "Ev Idx": 2492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836063.997, "dur": 2.066,
    "args": {
      "External id": 2494,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1422, "finished": true, "Ev Idx": 2493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836065.014, "dur": 0.507,
    "args": {
      "External id": 2495,"Record function id": 0, "finished": true, "Ev Idx": 2494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836066.629, "dur": 1.790,
    "args": {
      "External id": 2496,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1423, "finished": true, "Ev Idx": 2495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836067.653, "dur": 0.300,
    "args": {
      "External id": 2497,"Record function id": 0, "finished": true, "Ev Idx": 2496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564836071.209, "dur": 3.568,
    "args": {
      "External id": 2498,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1424, "finished": true, "Ev Idx": 2497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836073.545, "dur": 0.508,
    "args": {
      "External id": 2499,"Record function id": 0, "finished": true, "Ev Idx": 2498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836077.306, "dur": 20.784,
    "args": {
      "External id": 2500,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1425, "finished": true, "Ev Idx": 2499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564836080.080, "dur": 15.038,
    "args": {
      "External id": 2501,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1425, "finished": true, "Ev Idx": 2500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564836082.183, "dur": 2.774,
    "args": {
      "External id": 2502,"Record function id": 0, "finished": true, "Ev Idx": 2501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836083.026, "dur": 1.690,
    "args": {
      "External id": 2503,"Record function id": 0, "finished": true, "Ev Idx": 2502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564836085.572, "dur": 8.974,
    "args": {
      "External id": 2504,"Record function id": 0, "finished": true, "Ev Idx": 2503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564836095.639, "dur": 2.195,
    "args": {
      "External id": 2505,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1426, "finished": true, "Ev Idx": 2504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836101.516, "dur": 2.448,
    "args": {
      "External id": 2506,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1427, "finished": true, "Ev Idx": 2505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836103.005, "dur": 0.388,
    "args": {
      "External id": 2507,"Record function id": 0, "finished": true, "Ev Idx": 2506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836104.577, "dur": 1.470,
    "args": {
      "External id": 2508,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1428, "finished": true, "Ev Idx": 2507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836105.287, "dur": 0.287,
    "args": {
      "External id": 2509,"Record function id": 0, "finished": true, "Ev Idx": 2508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564836106.589, "dur": 1.439,
    "args": {
      "External id": 2510,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1429, "finished": true, "Ev Idx": 2509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836107.266, "dur": 0.284,
    "args": {
      "External id": 2511,"Record function id": 0, "finished": true, "Ev Idx": 2510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836108.745, "dur": 3.309,
    "args": {
      "External id": 2512,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1430, "finished": true, "Ev Idx": 2511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836111.096, "dur": 0.472,
    "args": {
      "External id": 2513,"Record function id": 0, "finished": true, "Ev Idx": 2512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836112.670, "dur": 2.218,
    "args": {
      "External id": 2514,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1431, "finished": true, "Ev Idx": 2513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836114.136, "dur": 0.262,
    "args": {
      "External id": 2515,"Record function id": 0, "finished": true, "Ev Idx": 2514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564836116.409, "dur": 1.865,
    "args": {
      "External id": 2516,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1432, "finished": true, "Ev Idx": 2515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836117.359, "dur": 0.477,
    "args": {
      "External id": 2517,"Record function id": 0, "finished": true, "Ev Idx": 2516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836119.685, "dur": 11.445,
    "args": {
      "External id": 2518,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1433, "finished": true, "Ev Idx": 2517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564836120.752, "dur": 8.765,
    "args": {
      "External id": 2519,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1433, "finished": true, "Ev Idx": 2518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564836121.635, "dur": 1.799,
    "args": {
      "External id": 2520,"Record function id": 0, "finished": true, "Ev Idx": 2519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836122.057, "dur": 1.264,
    "args": {
      "External id": 2521,"Record function id": 0, "finished": true, "Ev Idx": 2520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564836123.868, "dur": 5.298,
    "args": {
      "External id": 2522,"Record function id": 0, "finished": true, "Ev Idx": 2521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564836130.023, "dur": 0.939,
    "args": {
      "External id": 2523,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1434, "finished": true, "Ev Idx": 2522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564836133.180, "dur": 2.755,
    "args": {
      "External id": 2524,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1435, "finished": true, "Ev Idx": 2523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836135.024, "dur": 0.381,
    "args": {
      "External id": 2525,"Record function id": 0, "finished": true, "Ev Idx": 2524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564836137.516, "dur": 48.152,
    "args": {
      "External id": 2526,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1436, "finished": true, "Ev Idx": 2525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564836138.184, "dur": 2.559,
    "args": {
      "External id": 2527,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1436, "finished": true, "Ev Idx": 2526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836139.812, "dur": 0.480,
    "args": {
      "External id": 2528,"Record function id": 0, "finished": true, "Ev Idx": 2527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836141.089, "dur": 3.799,
    "args": {
      "External id": 2529,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1437, "finished": true, "Ev Idx": 2528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564836142.817, "dur": 1.896,
    "args": {
      "External id": 2530,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1437, "finished": true, "Ev Idx": 2529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564836145.357, "dur": 1.578,
    "args": {
      "External id": 2531,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1438, "finished": true, "Ev Idx": 2530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836146.045, "dur": 0.512,
    "args": {
      "External id": 2532,"Record function id": 0, "finished": true, "Ev Idx": 2531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836147.228, "dur": 2.169,
    "args": {
      "External id": 2533,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1439, "finished": true, "Ev Idx": 2532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564836148.246, "dur": 1.041,
    "args": {
      "External id": 2534,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1439, "finished": true, "Ev Idx": 2533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564836150.022, "dur": 33.536,
    "args": {
      "External id": 2535,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1440, "finished": true, "Ev Idx": 2534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836163.275, "dur": 1.143,
    "args": {
      "External id": 2536,"Record function id": 0, "finished": true, "Ev Idx": 2535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564836184.108, "dur": 1.177,
    "args": {
      "External id": 2537,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1441, "finished": true, "Ev Idx": 2536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564836188.973, "dur": 10.782,
    "args": {
      "External id": 2538,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1442, "finished": true, "Ev Idx": 2537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836202.925, "dur": 1.851,
    "args": {
      "External id": 2539,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1443, "finished": true, "Ev Idx": 2538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836203.654, "dur": 0.455,
    "args": {
      "External id": 2540,"Record function id": 0, "finished": true, "Ev Idx": 2539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836205.474, "dur": 1.431,
    "args": {
      "External id": 2541,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1443, "finished": true, "Ev Idx": 2540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836206.052, "dur": 0.458,
    "args": {
      "External id": 2542,"Record function id": 0, "finished": true, "Ev Idx": 2541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564836207.536, "dur": 1.009,
    "args": {
      "External id": 2543,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1443, "finished": true, "Ev Idx": 2542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836208.006, "dur": 0.228,
    "args": {
      "External id": 2544,"Record function id": 0, "finished": true, "Ev Idx": 2543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564836210.220, "dur": 7.975,
    "args": {
      "External id": 2545,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1443, "finished": true, "Ev Idx": 2544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564836224.161, "dur": 13.324,
    "args": {
      "External id": 2546,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1444, "finished": true, "Ev Idx": 2545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564836225.244, "dur": 0.415,
    "args": {
      "External id": 2547,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1444, "finished": true, "Ev Idx": 2546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564836226.409, "dur": 10.770,
    "args": {
      "External id": 2548,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1444, "finished": true, "Ev Idx": 2547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564836240.558, "dur": 0.212,
    "args": {
      "External id": 2549,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1445, "finished": true, "Ev Idx": 2548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564836251.410, "dur": 0.467,
    "args": {
      "External id": 2550,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1445, "finished": true, "Ev Idx": 2549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564836253.732, "dur": 73.658,
    "args": {
      "External id": 2551,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1445, "finished": true, "Ev Idx": 2550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564836254.285, "dur": 2.240,
    "args": {
      "External id": 2552,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1445, "finished": true, "Ev Idx": 2551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836255.560, "dur": 0.480,
    "args": {
      "External id": 2553,"Record function id": 0, "finished": true, "Ev Idx": 2552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836256.907, "dur": 2.846,
    "args": {
      "External id": 2554,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1446, "finished": true, "Ev Idx": 2553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564836257.324, "dur": 2.306,
    "args": {
      "External id": 2555,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1446, "finished": true, "Ev Idx": 2554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564836260.169, "dur": 2.499,
    "args": {
      "External id": 2556,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1447, "finished": true, "Ev Idx": 2555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836261.856, "dur": 0.438,
    "args": {
      "External id": 2557,"Record function id": 0, "finished": true, "Ev Idx": 2556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836262.970, "dur": 2.377,
    "args": {
      "External id": 2558,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1448, "finished": true, "Ev Idx": 2557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564836263.425, "dur": 1.839,
    "args": {
      "External id": 2559,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1448, "finished": true, "Ev Idx": 2558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564836265.754, "dur": 59.677,
    "args": {
      "External id": 2560,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1449, "finished": true, "Ev Idx": 2559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836308.401, "dur": 1.252,
    "args": {
      "External id": 2561,"Record function id": 0, "finished": true, "Ev Idx": 2560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564836326.026, "dur": 1.070,
    "args": {
      "External id": 2562,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1450, "finished": true, "Ev Idx": 2561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564836329.639, "dur": 1.789,
    "args": {
      "External id": 2563,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1451, "finished": true, "Ev Idx": 2562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836330.501, "dur": 0.371,
    "args": {
      "External id": 2564,"Record function id": 0, "finished": true, "Ev Idx": 2563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564836334.019, "dur": 11.326,
    "args": {
      "External id": 2565,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1452, "finished": true, "Ev Idx": 2564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564836334.686, "dur": 10.522,
    "args": {
      "External id": 2566,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1452, "finished": true, "Ev Idx": 2565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564836335.824, "dur": 1.831,
    "args": {
      "External id": 2567,"Record function id": 0, "finished": true, "Ev Idx": 2566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836336.178, "dur": 1.342,
    "args": {
      "External id": 2568,"Record function id": 0, "finished": true, "Ev Idx": 2567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564836338.010, "dur": 6.774,
    "args": {
      "External id": 2569,"Record function id": 0, "finished": true, "Ev Idx": 2568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836351.283, "dur": 1.808,
    "args": {
      "External id": 2570,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1453, "finished": true, "Ev Idx": 2569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564836351.746, "dur": 1.256,
    "args": {
      "External id": 2571,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1453, "finished": true, "Ev Idx": 2570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564836368.313, "dur": 36.848,
    "args": {
      "External id": 2572,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1454, "finished": true, "Ev Idx": 2571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564836368.914, "dur": 3.149,
    "args": {
      "External id": 2573,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1454, "finished": true, "Ev Idx": 2572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564836370.547, "dur": 0.961,
    "args": {
      "External id": 2574,"Record function id": 0, "finished": true, "Ev Idx": 2573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836371.008, "dur": 0.337,
    "args": {
      "External id": 2575,"Record function id": 0, "finished": true, "Ev Idx": 2574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564836372.503, "dur": 32.506,
    "args": {
      "External id": 2576,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1455, "finished": true, "Ev Idx": 2575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836372.942, "dur": 1.596,
    "args": {
      "External id": 2577,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1455, "finished": true, "Ev Idx": 2576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564836373.235, "dur": 1.213,
    "args": {
      "External id": 2578,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1455, "finished": true, "Ev Idx": 2577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564836375.133, "dur": 27.833,
    "args": {
      "External id": 2579,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1456, "finished": true, "Ev Idx": 2578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836387.436, "dur": 1.086,
    "args": {
      "External id": 2580,"Record function id": 0, "finished": true, "Ev Idx": 2579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564836403.614, "dur": 0.979,
    "args": {
      "External id": 2581,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1457, "finished": true, "Ev Idx": 2580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564836410.408, "dur": 9.744,
    "args": {
      "External id": 2582,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1458, "finished": true, "Ev Idx": 2581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564836429.948, "dur": 0.183,
    "args": {
      "External id": 2583,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1459, "finished": true, "Ev Idx": 2582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564836432.351, "dur": 12.436,
    "args": {
      "External id": 2584,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1459, "finished": true, "Ev Idx": 2583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564836434.471, "dur": 0.380,
    "args": {
      "External id": 2585,"Record function id": 0, "finished": true, "Ev Idx": 2584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564836435.349, "dur": 0.071,
    "args": {
      "External id": 2586,"Record function id": 0, "finished": true, "Ev Idx": 2585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564836447.740, "dur": 12.357,
    "args": {
      "External id": 2587,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1460, "finished": true, "Ev Idx": 2586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564836463.743, "dur": 8.940,
    "args": {
      "External id": 2588,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1461, "finished": true, "Ev Idx": 2587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564836474.705, "dur": 11.181,
    "args": {
      "External id": 2589,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1462, "finished": true, "Ev Idx": 2588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564836487.755, "dur": 8.909,
    "args": {
      "External id": 2590,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1463, "finished": true, "Ev Idx": 2589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564836499.270, "dur": 0.121,
    "args": {
      "External id": 2591,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1464, "finished": true, "Ev Idx": 2590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564836500.317, "dur": 10.053,
    "args": {
      "External id": 2592,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1464, "finished": true, "Ev Idx": 2591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564836524.637, "dur": 46.411,
    "args": {
      "External id": 2593,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1465, "finished": true, "Ev Idx": 2592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564836525.010, "dur": 3.120,
    "args": {
      "External id": 2594,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1465, "finished": true, "Ev Idx": 2593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564836526.172, "dur": 1.341,
    "args": {
      "External id": 2595,"Record function id": 0, "finished": true, "Ev Idx": 2594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836526.900, "dur": 0.487,
    "args": {
      "External id": 2596,"Record function id": 0, "finished": true, "Ev Idx": 2595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564836528.510, "dur": 42.406,
    "args": {
      "External id": 2597,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1466, "finished": true, "Ev Idx": 2596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836529.000, "dur": 2.815,
    "args": {
      "External id": 2598,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1466, "finished": true, "Ev Idx": 2597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564836529.328, "dur": 2.360,
    "args": {
      "External id": 2599,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1466, "finished": true, "Ev Idx": 2598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564836532.171, "dur": 36.705,
    "args": {
      "External id": 2600,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1467, "finished": true, "Ev Idx": 2599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836541.679, "dur": 1.191,
    "args": {
      "External id": 2601,"Record function id": 0, "finished": true, "Ev Idx": 2600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564836569.477, "dur": 1.113,
    "args": {
      "External id": 2602,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1468, "finished": true, "Ev Idx": 2601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564836579.123, "dur": 65.803,
    "args": {
      "External id": 2603,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1469, "finished": true, "Ev Idx": 2602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564836653.227, "dur": 46.924,
    "args": {
      "External id": 2604,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1470, "finished": true, "Ev Idx": 2603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564836653.625, "dur": 4.148,
    "args": {
      "External id": 2605,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1470, "finished": true, "Ev Idx": 2604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564836654.961, "dur": 1.668,
    "args": {
      "External id": 2606,"Record function id": 0, "finished": true, "Ev Idx": 2605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836655.702, "dur": 0.801,
    "args": {
      "External id": 2607,"Record function id": 0, "finished": true, "Ev Idx": 2606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564836658.155, "dur": 41.892,
    "args": {
      "External id": 2608,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1471, "finished": true, "Ev Idx": 2607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836658.999, "dur": 2.398,
    "args": {
      "External id": 2609,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1471, "finished": true, "Ev Idx": 2608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564836659.346, "dur": 1.912,
    "args": {
      "External id": 2610,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1471, "finished": true, "Ev Idx": 2609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564836661.775, "dur": 36.214,
    "args": {
      "External id": 2611,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1472, "finished": true, "Ev Idx": 2610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836674.341, "dur": 1.364,
    "args": {
      "External id": 2612,"Record function id": 0, "finished": true, "Ev Idx": 2611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564836698.597, "dur": 1.188,
    "args": {
      "External id": 2613,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1473, "finished": true, "Ev Idx": 2612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564836702.180, "dur": 41.787,
    "args": {
      "External id": 2614,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1474, "finished": true, "Ev Idx": 2613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564836750.913, "dur": 44.461,
    "args": {
      "External id": 2615,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1475, "finished": true, "Ev Idx": 2614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564836751.726, "dur": 2.723,
    "args": {
      "External id": 2616,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1475, "finished": true, "Ev Idx": 2615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564836752.606, "dur": 1.243,
    "args": {
      "External id": 2617,"Record function id": 0, "finished": true, "Ev Idx": 2616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836753.330, "dur": 0.383,
    "args": {
      "External id": 2618,"Record function id": 0, "finished": true, "Ev Idx": 2617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564836754.900, "dur": 40.358,
    "args": {
      "External id": 2619,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1476, "finished": true, "Ev Idx": 2618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836755.325, "dur": 1.694,
    "args": {
      "External id": 2620,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1476, "finished": true, "Ev Idx": 2619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564836755.839, "dur": 1.074,
    "args": {
      "External id": 2621,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1476, "finished": true, "Ev Idx": 2620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564836757.354, "dur": 36.055,
    "args": {
      "External id": 2622,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1477, "finished": true, "Ev Idx": 2621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836771.140, "dur": 1.119,
    "args": {
      "External id": 2623,"Record function id": 0, "finished": true, "Ev Idx": 2622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564836793.986, "dur": 0.937,
    "args": {
      "External id": 2624,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1478, "finished": true, "Ev Idx": 2623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564836798.057, "dur": 9.061,
    "args": {
      "External id": 2625,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1479, "finished": true, "Ev Idx": 2624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564836828.586, "dur": 0.184,
    "args": {
      "External id": 2626,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1480, "finished": true, "Ev Idx": 2625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564836830.854, "dur": 11.447,
    "args": {
      "External id": 2627,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1480, "finished": true, "Ev Idx": 2626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564836832.362, "dur": 0.806,
    "args": {
      "External id": 2628,"Record function id": 0, "finished": true, "Ev Idx": 2627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564836833.572, "dur": 0.088,
    "args": {
      "External id": 2629,"Record function id": 0, "finished": true, "Ev Idx": 2628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564836845.041, "dur": 12.485,
    "args": {
      "External id": 2630,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1481, "finished": true, "Ev Idx": 2629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564836861.390, "dur": 8.964,
    "args": {
      "External id": 2631,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1482, "finished": true, "Ev Idx": 2630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564836872.072, "dur": 9.309,
    "args": {
      "External id": 2632,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1483, "finished": true, "Ev Idx": 2631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564836883.061, "dur": 8.862,
    "args": {
      "External id": 2633,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1484, "finished": true, "Ev Idx": 2632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564836894.420, "dur": 0.095,
    "args": {
      "External id": 2634,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1485, "finished": true, "Ev Idx": 2633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564836895.400, "dur": 9.293,
    "args": {
      "External id": 2635,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1485, "finished": true, "Ev Idx": 2634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564836922.343, "dur": 37.089,
    "args": {
      "External id": 2636,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1486, "finished": true, "Ev Idx": 2635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564836922.800, "dur": 4.253,
    "args": {
      "External id": 2637,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1486, "finished": true, "Ev Idx": 2636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564836924.760, "dur": 1.401,
    "args": {
      "External id": 2638,"Record function id": 0, "finished": true, "Ev Idx": 2637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836925.521, "dur": 0.496,
    "args": {
      "External id": 2639,"Record function id": 0, "finished": true, "Ev Idx": 2638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564836927.364, "dur": 31.953,
    "args": {
      "External id": 2640,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1487, "finished": true, "Ev Idx": 2639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836928.072, "dur": 1.652,
    "args": {
      "External id": 2641,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1487, "finished": true, "Ev Idx": 2640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564836928.358, "dur": 1.232,
    "args": {
      "External id": 2642,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1487, "finished": true, "Ev Idx": 2641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564836930.056, "dur": 27.208,
    "args": {
      "External id": 2643,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1488, "finished": true, "Ev Idx": 2642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836942.105, "dur": 1.154,
    "args": {
      "External id": 2644,"Record function id": 0, "finished": true, "Ev Idx": 2643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564836957.859, "dur": 1.117,
    "args": {
      "External id": 2645,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1489, "finished": true, "Ev Idx": 2644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564836963.010, "dur": 1.450,
    "args": {
      "External id": 2646,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1490, "finished": true, "Ev Idx": 2645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564836966.645, "dur": 2.849,
    "args": {
      "External id": 2647,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1491, "finished": true, "Ev Idx": 2646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836968.022, "dur": 0.645,
    "args": {
      "External id": 2648,"Record function id": 0, "finished": true, "Ev Idx": 2647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564836976.593, "dur": 41.649,
    "args": {
      "External id": 2649,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1492, "finished": true, "Ev Idx": 2648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564836977.210, "dur": 2.228,
    "args": {
      "External id": 2650,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1492, "finished": true, "Ev Idx": 2649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564836978.154, "dur": 0.849,
    "args": {
      "External id": 2651,"Record function id": 0, "finished": true, "Ev Idx": 2650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564836978.629, "dur": 0.282,
    "args": {
      "External id": 2652,"Record function id": 0, "finished": true, "Ev Idx": 2651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564836979.795, "dur": 38.352,
    "args": {
      "External id": 2653,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1493, "finished": true, "Ev Idx": 2652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564836980.206, "dur": 1.253,
    "args": {
      "External id": 2654,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1493, "finished": true, "Ev Idx": 2653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564836980.458, "dur": 0.908,
    "args": {
      "External id": 2655,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1493, "finished": true, "Ev Idx": 2654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564836981.784, "dur": 34.307,
    "args": {
      "External id": 2656,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1494, "finished": true, "Ev Idx": 2655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564836992.269, "dur": 1.244,
    "args": {
      "External id": 2657,"Record function id": 0, "finished": true, "Ev Idx": 2656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564837016.673, "dur": 1.169,
    "args": {
      "External id": 2658,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1495, "finished": true, "Ev Idx": 2657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564837020.514, "dur": 1.174,
    "args": {
      "External id": 2659,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1496, "finished": true, "Ev Idx": 2658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564837023.519, "dur": 1.732,
    "args": {
      "External id": 2660,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1497, "finished": true, "Ev Idx": 2659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837024.292, "dur": 0.451,
    "args": {
      "External id": 2661,"Record function id": 0, "finished": true, "Ev Idx": 2660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564837031.410, "dur": 39.930,
    "args": {
      "External id": 2662,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1498, "finished": true, "Ev Idx": 2661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564837031.696, "dur": 2.584,
    "args": {
      "External id": 2663,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1498, "finished": true, "Ev Idx": 2662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564837032.836, "dur": 0.976,
    "args": {
      "External id": 2664,"Record function id": 0, "finished": true, "Ev Idx": 2663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837033.282, "dur": 0.427,
    "args": {
      "External id": 2665,"Record function id": 0, "finished": true, "Ev Idx": 2664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564837034.570, "dur": 36.674,
    "args": {
      "External id": 2666,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1499, "finished": true, "Ev Idx": 2665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837035.115, "dur": 1.478,
    "args": {
      "External id": 2667,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1499, "finished": true, "Ev Idx": 2666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564837035.479, "dur": 1.049,
    "args": {
      "External id": 2668,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1499, "finished": true, "Ev Idx": 2667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564837037.079, "dur": 32.283,
    "args": {
      "External id": 2669,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1500, "finished": true, "Ev Idx": 2668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564837049.842, "dur": 1.063,
    "args": {
      "External id": 2670,"Record function id": 0, "finished": true, "Ev Idx": 2669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564837070.024, "dur": 0.940,
    "args": {
      "External id": 2671,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1501, "finished": true, "Ev Idx": 2670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564837073.707, "dur": 1.104,
    "args": {
      "External id": 2672,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1502, "finished": true, "Ev Idx": 2671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564837076.240, "dur": 1.822,
    "args": {
      "External id": 2673,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1503, "finished": true, "Ev Idx": 2672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837077.111, "dur": 0.356,
    "args": {
      "External id": 2674,"Record function id": 0, "finished": true, "Ev Idx": 2673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564837081.119, "dur": 2.148,
    "args": {
      "External id": 2675,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1504, "finished": true, "Ev Idx": 2674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837082.401, "dur": 0.296,
    "args": {
      "External id": 2676,"Record function id": 0, "finished": true, "Ev Idx": 2675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564837084.341, "dur": 1.129,
    "args": {
      "External id": 2677,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1504, "finished": true, "Ev Idx": 2676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837084.846, "dur": 0.258,
    "args": {
      "External id": 2678,"Record function id": 0, "finished": true, "Ev Idx": 2677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564837086.290, "dur": 10.481,
    "args": {
      "External id": 2679,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1504, "finished": true, "Ev Idx": 2678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837102.075, "dur": 5.730,
    "args": {
      "External id": 2680,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1505, "finished": true, "Ev Idx": 2679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837105.433, "dur": 0.884,
    "args": {
      "External id": 2681,"Record function id": 0, "finished": true, "Ev Idx": 2680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837110.068, "dur": 1.761,
    "args": {
      "External id": 2682,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1506, "finished": true, "Ev Idx": 2681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837110.881, "dur": 0.354,
    "args": {
      "External id": 2683,"Record function id": 0, "finished": true, "Ev Idx": 2682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564837113.855, "dur": 10.759,
    "args": {
      "External id": 2684,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1507, "finished": true, "Ev Idx": 2683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564837129.030, "dur": 17.762,
    "args": {
      "External id": 2685,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1508, "finished": true, "Ev Idx": 2684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564837149.481, "dur": 9.663,
    "args": {
      "External id": 2686,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1509, "finished": true, "Ev Idx": 2685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564837160.717, "dur": 9.565,
    "args": {
      "External id": 2687,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1510, "finished": true, "Ev Idx": 2686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564837172.265, "dur": 6.907,
    "args": {
      "External id": 2688,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1511, "finished": true, "Ev Idx": 2687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837181.540, "dur": 2.176,
    "args": {
      "External id": 2689,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1512, "finished": true, "Ev Idx": 2688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837182.603, "dur": 0.438,
    "args": {
      "External id": 2690,"Record function id": 0, "finished": true, "Ev Idx": 2689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837185.224, "dur": 1.834,
    "args": {
      "External id": 2691,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1513, "finished": true, "Ev Idx": 2690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837186.118, "dur": 0.421,
    "args": {
      "External id": 2692,"Record function id": 0, "finished": true, "Ev Idx": 2691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564837188.135, "dur": 7.436,
    "args": {
      "External id": 2693,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1514, "finished": true, "Ev Idx": 2692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564837197.687, "dur": 9.067,
    "args": {
      "External id": 2694,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1515, "finished": true, "Ev Idx": 2693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564837208.742, "dur": 7.774,
    "args": {
      "External id": 2695,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1516, "finished": true, "Ev Idx": 2694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564837218.274, "dur": 7.585,
    "args": {
      "External id": 2696,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1517, "finished": true, "Ev Idx": 2695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837249.980, "dur": 2.662,
    "args": {
      "External id": 2697,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1518, "finished": true, "Ev Idx": 2696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837251.375, "dur": 0.486,
    "args": {
      "External id": 2698,"Record function id": 0, "finished": true, "Ev Idx": 2697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837253.325, "dur": 1.582,
    "args": {
      "External id": 2699,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1519, "finished": true, "Ev Idx": 2698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837254.064, "dur": 0.359,
    "args": {
      "External id": 2700,"Record function id": 0, "finished": true, "Ev Idx": 2699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564837255.574, "dur": 3.260,
    "args": {
      "External id": 2701,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1520, "finished": true, "Ev Idx": 2700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837257.454, "dur": 0.646,
    "args": {
      "External id": 2702,"Record function id": 0, "finished": true, "Ev Idx": 2701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837259.550, "dur": 3.014,
    "args": {
      "External id": 2703,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1521, "finished": true, "Ev Idx": 2702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837261.404, "dur": 0.563,
    "args": {
      "External id": 2704,"Record function id": 0, "finished": true, "Ev Idx": 2703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837263.153, "dur": 1.834,
    "args": {
      "External id": 2705,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1522, "finished": true, "Ev Idx": 2704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837264.201, "dur": 0.329,
    "args": {
      "External id": 2706,"Record function id": 0, "finished": true, "Ev Idx": 2705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564837267.833, "dur": 2.754,
    "args": {
      "External id": 2707,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1523, "finished": true, "Ev Idx": 2706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837269.483, "dur": 0.394,
    "args": {
      "External id": 2708,"Record function id": 0, "finished": true, "Ev Idx": 2707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837273.066, "dur": 19.906,
    "args": {
      "External id": 2709,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1524, "finished": true, "Ev Idx": 2708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564837275.635, "dur": 15.108,
    "args": {
      "External id": 2710,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1524, "finished": true, "Ev Idx": 2709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564837277.211, "dur": 2.676,
    "args": {
      "External id": 2711,"Record function id": 0, "finished": true, "Ev Idx": 2710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564837278.031, "dur": 1.637,
    "args": {
      "External id": 2712,"Record function id": 0, "finished": true, "Ev Idx": 2711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564837280.605, "dur": 9.545,
    "args": {
      "External id": 2713,"Record function id": 0, "finished": true, "Ev Idx": 2712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564837291.336, "dur": 1.375,
    "args": {
      "External id": 2714,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1525, "finished": true, "Ev Idx": 2713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837296.318, "dur": 2.727,
    "args": {
      "External id": 2715,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1526, "finished": true, "Ev Idx": 2714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837298.072, "dur": 0.376,
    "args": {
      "External id": 2716,"Record function id": 0, "finished": true, "Ev Idx": 2715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837299.629, "dur": 2.058,
    "args": {
      "External id": 2717,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1527, "finished": true, "Ev Idx": 2716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837300.878, "dur": 0.307,
    "args": {
      "External id": 2718,"Record function id": 0, "finished": true, "Ev Idx": 2717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564837302.230, "dur": 1.457,
    "args": {
      "External id": 2719,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1528, "finished": true, "Ev Idx": 2718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837302.936, "dur": 0.304,
    "args": {
      "External id": 2720,"Record function id": 0, "finished": true, "Ev Idx": 2719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837304.292, "dur": 1.753,
    "args": {
      "External id": 2721,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1529, "finished": true, "Ev Idx": 2720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837305.213, "dur": 0.325,
    "args": {
      "External id": 2722,"Record function id": 0, "finished": true, "Ev Idx": 2721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837306.687, "dur": 1.465,
    "args": {
      "External id": 2723,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1530, "finished": true, "Ev Idx": 2722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837307.375, "dur": 0.280,
    "args": {
      "External id": 2724,"Record function id": 0, "finished": true, "Ev Idx": 2723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564837309.821, "dur": 1.793,
    "args": {
      "External id": 2725,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1531, "finished": true, "Ev Idx": 2724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837310.833, "dur": 0.341,
    "args": {
      "External id": 2726,"Record function id": 0, "finished": true, "Ev Idx": 2725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837313.109, "dur": 12.387,
    "args": {
      "External id": 2727,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1532, "finished": true, "Ev Idx": 2726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564837314.152, "dur": 9.892,
    "args": {
      "External id": 2728,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1532, "finished": true, "Ev Idx": 2727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564837315.541, "dur": 1.630,
    "args": {
      "External id": 2729,"Record function id": 0, "finished": true, "Ev Idx": 2728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564837315.885, "dur": 1.173,
    "args": {
      "External id": 2730,"Record function id": 0, "finished": true, "Ev Idx": 2729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564837317.527, "dur": 6.199,
    "args": {
      "External id": 2731,"Record function id": 0, "finished": true, "Ev Idx": 2730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564837324.439, "dur": 0.883,
    "args": {
      "External id": 2732,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1533, "finished": true, "Ev Idx": 2731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564837327.425, "dur": 1.698,
    "args": {
      "External id": 2733,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1534, "finished": true, "Ev Idx": 2732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837328.238, "dur": 0.412,
    "args": {
      "External id": 2734,"Record function id": 0, "finished": true, "Ev Idx": 2733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564837330.597, "dur": 42.945,
    "args": {
      "External id": 2735,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1535, "finished": true, "Ev Idx": 2734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564837331.140, "dur": 1.550,
    "args": {
      "External id": 2736,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1535, "finished": true, "Ev Idx": 2735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837331.935, "dur": 0.321,
    "args": {
      "External id": 2737,"Record function id": 0, "finished": true, "Ev Idx": 2736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837333.037, "dur": 3.579,
    "args": {
      "External id": 2738,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1536, "finished": true, "Ev Idx": 2737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564837334.717, "dur": 1.736,
    "args": {
      "External id": 2739,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1536, "finished": true, "Ev Idx": 2738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564837337.105, "dur": 2.602,
    "args": {
      "External id": 2740,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1537, "finished": true, "Ev Idx": 2739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837338.918, "dur": 0.449,
    "args": {
      "External id": 2741,"Record function id": 0, "finished": true, "Ev Idx": 2740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837339.997, "dur": 2.644,
    "args": {
      "External id": 2742,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1538, "finished": true, "Ev Idx": 2741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564837340.813, "dur": 1.724,
    "args": {
      "External id": 2743,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1538, "finished": true, "Ev Idx": 2742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564837343.255, "dur": 28.199,
    "args": {
      "External id": 2744,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1539, "finished": true, "Ev Idx": 2743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564837355.102, "dur": 0.976,
    "args": {
      "External id": 2745,"Record function id": 0, "finished": true, "Ev Idx": 2744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564837372.008, "dur": 1.201,
    "args": {
      "External id": 2746,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1540, "finished": true, "Ev Idx": 2745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564837376.611, "dur": 9.176,
    "args": {
      "External id": 2747,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1541, "finished": true, "Ev Idx": 2746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837388.896, "dur": 1.752,
    "args": {
      "External id": 2748,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1542, "finished": true, "Ev Idx": 2747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837389.628, "dur": 0.412,
    "args": {
      "External id": 2749,"Record function id": 0, "finished": true, "Ev Idx": 2748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837391.273, "dur": 1.368,
    "args": {
      "External id": 2750,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1542, "finished": true, "Ev Idx": 2749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837392.007, "dur": 0.296,
    "args": {
      "External id": 2751,"Record function id": 0, "finished": true, "Ev Idx": 2750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564837393.632, "dur": 1.158,
    "args": {
      "External id": 2752,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1542, "finished": true, "Ev Idx": 2751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837394.232, "dur": 0.226,
    "args": {
      "External id": 2753,"Record function id": 0, "finished": true, "Ev Idx": 2752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564837396.501, "dur": 8.155,
    "args": {
      "External id": 2754,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1542, "finished": true, "Ev Idx": 2753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564837409.757, "dur": 13.142,
    "args": {
      "External id": 2755,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1543, "finished": true, "Ev Idx": 2754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564837410.699, "dur": 0.368,
    "args": {
      "External id": 2756,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1543, "finished": true, "Ev Idx": 2755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564837411.740, "dur": 10.860,
    "args": {
      "External id": 2757,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1543, "finished": true, "Ev Idx": 2756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564837425.491, "dur": 0.181,
    "args": {
      "External id": 2758,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1544, "finished": true, "Ev Idx": 2757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564837437.327, "dur": 0.374,
    "args": {
      "External id": 2759,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1544, "finished": true, "Ev Idx": 2758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564837439.349, "dur": 42.035,
    "args": {
      "External id": 2760,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1544, "finished": true, "Ev Idx": 2759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564837440.074, "dur": 2.806,
    "args": {
      "External id": 2761,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1544, "finished": true, "Ev Idx": 2760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837442.042, "dur": 0.335,
    "args": {
      "External id": 2762,"Record function id": 0, "finished": true, "Ev Idx": 2761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837443.364, "dur": 2.786,
    "args": {
      "External id": 2763,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1545, "finished": true, "Ev Idx": 2762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564837443.693, "dur": 2.343,
    "args": {
      "External id": 2764,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1545, "finished": true, "Ev Idx": 2763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564837446.705, "dur": 1.346,
    "args": {
      "External id": 2765,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1546, "finished": true, "Ev Idx": 2764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837447.462, "dur": 0.238,
    "args": {
      "External id": 2766,"Record function id": 0, "finished": true, "Ev Idx": 2765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837448.313, "dur": 1.496,
    "args": {
      "External id": 2767,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1547, "finished": true, "Ev Idx": 2766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564837448.620, "dur": 1.118,
    "args": {
      "External id": 2768,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1547, "finished": true, "Ev Idx": 2767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564837450.199, "dur": 29.233,
    "args": {
      "External id": 2769,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1548, "finished": true, "Ev Idx": 2768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564837462.188, "dur": 1.217,
    "args": {
      "External id": 2770,"Record function id": 0, "finished": true, "Ev Idx": 2769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564837480.042, "dur": 1.066,
    "args": {
      "External id": 2771,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1549, "finished": true, "Ev Idx": 2770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564837483.267, "dur": 1.901,
    "args": {
      "External id": 2772,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1550, "finished": true, "Ev Idx": 2771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837484.250, "dur": 0.384,
    "args": {
      "External id": 2773,"Record function id": 0, "finished": true, "Ev Idx": 2772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564837487.803, "dur": 11.136,
    "args": {
      "External id": 2774,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1551, "finished": true, "Ev Idx": 2773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564837488.409, "dur": 10.386,
    "args": {
      "External id": 2775,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1551, "finished": true, "Ev Idx": 2774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564837489.397, "dur": 1.683,
    "args": {
      "External id": 2776,"Record function id": 0, "finished": true, "Ev Idx": 2775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564837489.769, "dur": 1.173,
    "args": {
      "External id": 2777,"Record function id": 0, "finished": true, "Ev Idx": 2776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564837491.566, "dur": 6.821,
    "args": {
      "External id": 2778,"Record function id": 0, "finished": true, "Ev Idx": 2777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837504.274, "dur": 1.631,
    "args": {
      "External id": 2779,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1552, "finished": true, "Ev Idx": 2778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564837504.631, "dur": 1.181,
    "args": {
      "External id": 2780,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1552, "finished": true, "Ev Idx": 2779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564837518.688, "dur": 36.030,
    "args": {
      "External id": 2781,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1553, "finished": true, "Ev Idx": 2780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564837519.147, "dur": 2.732,
    "args": {
      "External id": 2782,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1553, "finished": true, "Ev Idx": 2781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564837520.402, "dur": 0.896,
    "args": {
      "External id": 2783,"Record function id": 0, "finished": true, "Ev Idx": 2782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837520.860, "dur": 0.337,
    "args": {
      "External id": 2784,"Record function id": 0, "finished": true, "Ev Idx": 2783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564837522.253, "dur": 32.321,
    "args": {
      "External id": 2785,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1554, "finished": true, "Ev Idx": 2784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837522.675, "dur": 2.196,
    "args": {
      "External id": 2786,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1554, "finished": true, "Ev Idx": 2785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564837523.147, "dur": 1.625,
    "args": {
      "External id": 2787,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1554, "finished": true, "Ev Idx": 2786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564837525.305, "dur": 27.274,
    "args": {
      "External id": 2788,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1555, "finished": true, "Ev Idx": 2787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564837537.488, "dur": 1.034,
    "args": {
      "External id": 2789,"Record function id": 0, "finished": true, "Ev Idx": 2788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564837553.163, "dur": 1.055,
    "args": {
      "External id": 2790,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1556, "finished": true, "Ev Idx": 2789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564837558.964, "dur": 9.085,
    "args": {
      "External id": 2791,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1557, "finished": true, "Ev Idx": 2790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564837576.349, "dur": 0.173,
    "args": {
      "External id": 2792,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1558, "finished": true, "Ev Idx": 2791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564837578.318, "dur": 11.580,
    "args": {
      "External id": 2793,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1558, "finished": true, "Ev Idx": 2792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564837579.969, "dur": 0.567,
    "args": {
      "External id": 2794,"Record function id": 0, "finished": true, "Ev Idx": 2793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564837580.988, "dur": 0.199,
    "args": {
      "External id": 2795,"Record function id": 0, "finished": true, "Ev Idx": 2794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564837592.565, "dur": 12.826,
    "args": {
      "External id": 2796,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1559, "finished": true, "Ev Idx": 2795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564837608.751, "dur": 8.944,
    "args": {
      "External id": 2797,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1560, "finished": true, "Ev Idx": 2796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564837619.663, "dur": 7.955,
    "args": {
      "External id": 2798,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1561, "finished": true, "Ev Idx": 2797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564837629.310, "dur": 8.490,
    "args": {
      "External id": 2799,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1562, "finished": true, "Ev Idx": 2798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564837640.486, "dur": 0.105,
    "args": {
      "External id": 2800,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1563, "finished": true, "Ev Idx": 2799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564837641.673, "dur": 62.422,
    "args": {
      "External id": 2801,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1563, "finished": true, "Ev Idx": 2800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564837719.761, "dur": 51.173,
    "args": {
      "External id": 2802,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1564, "finished": true, "Ev Idx": 2801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564837720.504, "dur": 4.234,
    "args": {
      "External id": 2803,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1564, "finished": true, "Ev Idx": 2802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564837722.132, "dur": 1.857,
    "args": {
      "External id": 2804,"Record function id": 0, "finished": true, "Ev Idx": 2803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837723.150, "dur": 0.678,
    "args": {
      "External id": 2805,"Record function id": 0, "finished": true, "Ev Idx": 2804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564837725.269, "dur": 45.534,
    "args": {
      "External id": 2806,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1565, "finished": true, "Ev Idx": 2805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837726.065, "dur": 2.800,
    "args": {
      "External id": 2807,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1565, "finished": true, "Ev Idx": 2806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564837726.583, "dur": 2.146,
    "args": {
      "External id": 2808,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1565, "finished": true, "Ev Idx": 2807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564837729.330, "dur": 39.301,
    "args": {
      "External id": 2809,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1566, "finished": true, "Ev Idx": 2808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564837741.148, "dur": 1.253,
    "args": {
      "External id": 2810,"Record function id": 0, "finished": true, "Ev Idx": 2809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564837769.278, "dur": 1.142,
    "args": {
      "External id": 2811,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1567, "finished": true, "Ev Idx": 2810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564837779.146, "dur": 10.545,
    "args": {
      "External id": 2812,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1568, "finished": true, "Ev Idx": 2811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564837796.707, "dur": 85.345,
    "args": {
      "External id": 2813,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1569, "finished": true, "Ev Idx": 2812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564837797.131, "dur": 2.851,
    "args": {
      "External id": 2814,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1569, "finished": true, "Ev Idx": 2813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564837798.362, "dur": 0.958,
    "args": {
      "External id": 2815,"Record function id": 0, "finished": true, "Ev Idx": 2814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837798.763, "dur": 0.448,
    "args": {
      "External id": 2816,"Record function id": 0, "finished": true, "Ev Idx": 2815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564837800.348, "dur": 81.619,
    "args": {
      "External id": 2817,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1570, "finished": true, "Ev Idx": 2816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837800.965, "dur": 1.957,
    "args": {
      "External id": 2818,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1570, "finished": true, "Ev Idx": 2817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564837801.384, "dur": 1.438,
    "args": {
      "External id": 2819,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1570, "finished": true, "Ev Idx": 2818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564837803.250, "dur": 76.635,
    "args": {
      "External id": 2820,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1571, "finished": true, "Ev Idx": 2819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564837849.885, "dur": 1.200,
    "args": {
      "External id": 2821,"Record function id": 0, "finished": true, "Ev Idx": 2820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564837880.483, "dur": 1.134,
    "args": {
      "External id": 2822,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1572, "finished": true, "Ev Idx": 2821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564837884.221, "dur": 8.717,
    "args": {
      "External id": 2823,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1573, "finished": true, "Ev Idx": 2822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564837899.373, "dur": 37.027,
    "args": {
      "External id": 2824,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1574, "finished": true, "Ev Idx": 2823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564837899.868, "dur": 3.076,
    "args": {
      "External id": 2825,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1574, "finished": true, "Ev Idx": 2824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564837901.283, "dur": 1.018,
    "args": {
      "External id": 2826,"Record function id": 0, "finished": true, "Ev Idx": 2825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564837901.759, "dur": 0.438,
    "args": {
      "External id": 2827,"Record function id": 0, "finished": true, "Ev Idx": 2826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564837903.225, "dur": 33.053,
    "args": {
      "External id": 2828,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1575, "finished": true, "Ev Idx": 2827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564837903.652, "dur": 1.702,
    "args": {
      "External id": 2829,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1575, "finished": true, "Ev Idx": 2828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564837903.945, "dur": 1.303,
    "args": {
      "External id": 2830,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1575, "finished": true, "Ev Idx": 2829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564837905.666, "dur": 28.728,
    "args": {
      "External id": 2831,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1576, "finished": true, "Ev Idx": 2830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564837913.599, "dur": 1.109,
    "args": {
      "External id": 2832,"Record function id": 0, "finished": true, "Ev Idx": 2831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564837934.970, "dur": 1.001,
    "args": {
      "External id": 2833,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1577, "finished": true, "Ev Idx": 2832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564837939.059, "dur": 9.031,
    "args": {
      "External id": 2834,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1578, "finished": true, "Ev Idx": 2833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564837968.166, "dur": 0.253,
    "args": {
      "External id": 2835,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1579, "finished": true, "Ev Idx": 2834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564837970.427, "dur": 10.095,
    "args": {
      "External id": 2836,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1579, "finished": true, "Ev Idx": 2835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564837972.027, "dur": 0.458,
    "args": {
      "External id": 2837,"Record function id": 0, "finished": true, "Ev Idx": 2836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564837972.778, "dur": 0.072,
    "args": {
      "External id": 2838,"Record function id": 0, "finished": true, "Ev Idx": 2837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564837983.188, "dur": 11.683,
    "args": {
      "External id": 2839,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1580, "finished": true, "Ev Idx": 2838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564837998.692, "dur": 7.492,
    "args": {
      "External id": 2840,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1581, "finished": true, "Ev Idx": 2839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564838007.966, "dur": 8.214,
    "args": {
      "External id": 2841,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1582, "finished": true, "Ev Idx": 2840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564838017.800, "dur": 8.169,
    "args": {
      "External id": 2842,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1583, "finished": true, "Ev Idx": 2841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564838028.323, "dur": 0.124,
    "args": {
      "External id": 2843,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1584, "finished": true, "Ev Idx": 2842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564838029.306, "dur": 8.605,
    "args": {
      "External id": 2844,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1584, "finished": true, "Ev Idx": 2843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564838057.808, "dur": 104.325,
    "args": {
      "External id": 2845,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1585, "finished": true, "Ev Idx": 2844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564838058.202, "dur": 2.792,
    "args": {
      "External id": 2846,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1585, "finished": true, "Ev Idx": 2845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564838059.134, "dur": 1.223,
    "args": {
      "External id": 2847,"Record function id": 0, "finished": true, "Ev Idx": 2846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838059.690, "dur": 0.537,
    "args": {
      "External id": 2848,"Record function id": 0, "finished": true, "Ev Idx": 2847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564838061.407, "dur": 100.587,
    "args": {
      "External id": 2849,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1586, "finished": true, "Ev Idx": 2848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838061.903, "dur": 1.749,
    "args": {
      "External id": 2850,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1586, "finished": true, "Ev Idx": 2849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838062.356, "dur": 1.168,
    "args": {
      "External id": 2851,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1586, "finished": true, "Ev Idx": 2850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564838063.990, "dur": 94.703,
    "args": {
      "External id": 2852,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1587, "finished": true, "Ev Idx": 2851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564838135.777, "dur": 1.952,
    "args": {
      "External id": 2853,"Record function id": 0, "finished": true, "Ev Idx": 2852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564838159.514, "dur": 1.968,
    "args": {
      "External id": 2854,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1588, "finished": true, "Ev Idx": 2853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838167.352, "dur": 1.901,
    "args": {
      "External id": 2855,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1589, "finished": true, "Ev Idx": 2854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564838171.877, "dur": 3.447,
    "args": {
      "External id": 2856,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1590, "finished": true, "Ev Idx": 2855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838173.472, "dur": 0.877,
    "args": {
      "External id": 2857,"Record function id": 0, "finished": true, "Ev Idx": 2856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564838184.342, "dur": 45.644,
    "args": {
      "External id": 2858,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1591, "finished": true, "Ev Idx": 2857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564838185.020, "dur": 2.911,
    "args": {
      "External id": 2859,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1591, "finished": true, "Ev Idx": 2858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564838186.206, "dur": 1.265,
    "args": {
      "External id": 2860,"Record function id": 0, "finished": true, "Ev Idx": 2859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838186.938, "dur": 0.427,
    "args": {
      "External id": 2861,"Record function id": 0, "finished": true, "Ev Idx": 2860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564838188.380, "dur": 41.519,
    "args": {
      "External id": 2862,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1592, "finished": true, "Ev Idx": 2861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838189.034, "dur": 1.519,
    "args": {
      "External id": 2863,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1592, "finished": true, "Ev Idx": 2862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838189.365, "dur": 1.076,
    "args": {
      "External id": 2864,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1592, "finished": true, "Ev Idx": 2863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564838190.905, "dur": 36.951,
    "args": {
      "External id": 2865,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1593, "finished": true, "Ev Idx": 2864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564838202.879, "dur": 1.387,
    "args": {
      "External id": 2866,"Record function id": 0, "finished": true, "Ev Idx": 2865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564838228.442, "dur": 1.090,
    "args": {
      "External id": 2867,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1594, "finished": true, "Ev Idx": 2866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838232.554, "dur": 1.207,
    "args": {
      "External id": 2868,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1595, "finished": true, "Ev Idx": 2867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564838235.333, "dur": 1.940,
    "args": {
      "External id": 2869,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1596, "finished": true, "Ev Idx": 2868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838236.361, "dur": 0.409,
    "args": {
      "External id": 2870,"Record function id": 0, "finished": true, "Ev Idx": 2869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564838245.216, "dur": 39.625,
    "args": {
      "External id": 2871,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1597, "finished": true, "Ev Idx": 2870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564838245.528, "dur": 2.450,
    "args": {
      "External id": 2872,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1597, "finished": true, "Ev Idx": 2871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564838246.332, "dur": 1.219,
    "args": {
      "External id": 2873,"Record function id": 0, "finished": true, "Ev Idx": 2872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838247.019, "dur": 0.432,
    "args": {
      "External id": 2874,"Record function id": 0, "finished": true, "Ev Idx": 2873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564838248.430, "dur": 36.314,
    "args": {
      "External id": 2875,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1598, "finished": true, "Ev Idx": 2874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838248.836, "dur": 2.422,
    "args": {
      "External id": 2876,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1598, "finished": true, "Ev Idx": 2875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838249.194, "dur": 1.966,
    "args": {
      "External id": 2877,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1598, "finished": true, "Ev Idx": 2876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564838251.601, "dur": 31.314,
    "args": {
      "External id": 2878,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1599, "finished": true, "Ev Idx": 2877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564838262.339, "dur": 1.082,
    "args": {
      "External id": 2879,"Record function id": 0, "finished": true, "Ev Idx": 2878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564838283.429, "dur": 1.014,
    "args": {
      "External id": 2880,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1600, "finished": true, "Ev Idx": 2879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838287.429, "dur": 1.099,
    "args": {
      "External id": 2881,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1601, "finished": true, "Ev Idx": 2880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564838289.937, "dur": 1.581,
    "args": {
      "External id": 2882,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1602, "finished": true, "Ev Idx": 2881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838290.638, "dur": 0.386,
    "args": {
      "External id": 2883,"Record function id": 0, "finished": true, "Ev Idx": 2882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564838294.630, "dur": 2.425,
    "args": {
      "External id": 2884,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1603, "finished": true, "Ev Idx": 2883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838296.163, "dur": 0.332,
    "args": {
      "External id": 2885,"Record function id": 0, "finished": true, "Ev Idx": 2884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564838298.302, "dur": 1.229,
    "args": {
      "External id": 2886,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1603, "finished": true, "Ev Idx": 2885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838298.949, "dur": 0.258,
    "args": {
      "External id": 2887,"Record function id": 0, "finished": true, "Ev Idx": 2886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564838300.664, "dur": 13.629,
    "args": {
      "External id": 2888,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1603, "finished": true, "Ev Idx": 2887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838320.631, "dur": 5.685,
    "args": {
      "External id": 2889,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1604, "finished": true, "Ev Idx": 2888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838323.841, "dur": 0.864,
    "args": {
      "External id": 2890,"Record function id": 0, "finished": true, "Ev Idx": 2889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838328.901, "dur": 1.968,
    "args": {
      "External id": 2891,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1605, "finished": true, "Ev Idx": 2890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838329.913, "dur": 0.385,
    "args": {
      "External id": 2892,"Record function id": 0, "finished": true, "Ev Idx": 2891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564838332.844, "dur": 11.907,
    "args": {
      "External id": 2893,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1606, "finished": true, "Ev Idx": 2892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564838349.386, "dur": 19.204,
    "args": {
      "External id": 2894,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1607, "finished": true, "Ev Idx": 2893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564838371.530, "dur": 7.879,
    "args": {
      "External id": 2895,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1608, "finished": true, "Ev Idx": 2894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564838381.408, "dur": 10.341,
    "args": {
      "External id": 2896,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1609, "finished": true, "Ev Idx": 2895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564838393.869, "dur": 7.776,
    "args": {
      "External id": 2897,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1610, "finished": true, "Ev Idx": 2896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838404.122, "dur": 2.135,
    "args": {
      "External id": 2898,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1611, "finished": true, "Ev Idx": 2897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838405.137, "dur": 0.412,
    "args": {
      "External id": 2899,"Record function id": 0, "finished": true, "Ev Idx": 2898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838407.831, "dur": 1.826,
    "args": {
      "External id": 2900,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1612, "finished": true, "Ev Idx": 2899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838408.799, "dur": 0.329,
    "args": {
      "External id": 2901,"Record function id": 0, "finished": true, "Ev Idx": 2900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564838410.799, "dur": 9.458,
    "args": {
      "External id": 2902,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1613, "finished": true, "Ev Idx": 2901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564838422.084, "dur": 9.773,
    "args": {
      "External id": 2903,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1614, "finished": true, "Ev Idx": 2902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564838434.248, "dur": 7.974,
    "args": {
      "External id": 2904,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1615, "finished": true, "Ev Idx": 2903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564838443.846, "dur": 6.848,
    "args": {
      "External id": 2905,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1616, "finished": true, "Ev Idx": 2904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838475.576, "dur": 2.527,
    "args": {
      "External id": 2906,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1617, "finished": true, "Ev Idx": 2905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838476.886, "dur": 0.516,
    "args": {
      "External id": 2907,"Record function id": 0, "finished": true, "Ev Idx": 2906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838478.699, "dur": 1.697,
    "args": {
      "External id": 2908,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1618, "finished": true, "Ev Idx": 2907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838479.526, "dur": 0.414,
    "args": {
      "External id": 2909,"Record function id": 0, "finished": true, "Ev Idx": 2908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564838481.162, "dur": 3.321,
    "args": {
      "External id": 2910,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1619, "finished": true, "Ev Idx": 2909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838482.983, "dur": 0.791,
    "args": {
      "External id": 2911,"Record function id": 0, "finished": true, "Ev Idx": 2910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838485.348, "dur": 3.098,
    "args": {
      "External id": 2912,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1620, "finished": true, "Ev Idx": 2911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838487.336, "dur": 0.537,
    "args": {
      "External id": 2913,"Record function id": 0, "finished": true, "Ev Idx": 2912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838489.391, "dur": 2.262,
    "args": {
      "External id": 2914,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1621, "finished": true, "Ev Idx": 2913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838490.560, "dur": 0.628,
    "args": {
      "External id": 2915,"Record function id": 0, "finished": true, "Ev Idx": 2914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564838494.904, "dur": 3.011,
    "args": {
      "External id": 2916,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1622, "finished": true, "Ev Idx": 2915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838496.796, "dur": 0.386,
    "args": {
      "External id": 2917,"Record function id": 0, "finished": true, "Ev Idx": 2916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838500.595, "dur": 20.795,
    "args": {
      "External id": 2918,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1623, "finished": true, "Ev Idx": 2917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564838503.333, "dur": 15.971,
    "args": {
      "External id": 2919,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1623, "finished": true, "Ev Idx": 2918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564838504.940, "dur": 2.810,
    "args": {
      "External id": 2920,"Record function id": 0, "finished": true, "Ev Idx": 2919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564838505.738, "dur": 1.734,
    "args": {
      "External id": 2921,"Record function id": 0, "finished": true, "Ev Idx": 2920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564838508.536, "dur": 10.099,
    "args": {
      "External id": 2922,"Record function id": 0, "finished": true, "Ev Idx": 2921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564838519.840, "dur": 1.251,
    "args": {
      "External id": 2923,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1624, "finished": true, "Ev Idx": 2922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838524.756, "dur": 2.706,
    "args": {
      "External id": 2924,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1625, "finished": true, "Ev Idx": 2923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838526.368, "dur": 0.524,
    "args": {
      "External id": 2925,"Record function id": 0, "finished": true, "Ev Idx": 2924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838528.041, "dur": 1.974,
    "args": {
      "External id": 2926,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1626, "finished": true, "Ev Idx": 2925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838529.308, "dur": 0.285,
    "args": {
      "External id": 2927,"Record function id": 0, "finished": true, "Ev Idx": 2926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564838530.584, "dur": 1.483,
    "args": {
      "External id": 2928,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1627, "finished": true, "Ev Idx": 2927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838531.276, "dur": 0.356,
    "args": {
      "External id": 2929,"Record function id": 0, "finished": true, "Ev Idx": 2928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838532.681, "dur": 1.691,
    "args": {
      "External id": 2930,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1628, "finished": true, "Ev Idx": 2929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838533.585, "dur": 0.328,
    "args": {
      "External id": 2931,"Record function id": 0, "finished": true, "Ev Idx": 2930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838534.967, "dur": 1.466,
    "args": {
      "External id": 2932,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1629, "finished": true, "Ev Idx": 2931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838535.698, "dur": 0.287,
    "args": {
      "External id": 2933,"Record function id": 0, "finished": true, "Ev Idx": 2932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564838538.067, "dur": 1.717,
    "args": {
      "External id": 2934,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1630, "finished": true, "Ev Idx": 2933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838539.064, "dur": 0.318,
    "args": {
      "External id": 2935,"Record function id": 0, "finished": true, "Ev Idx": 2934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838541.197, "dur": 13.794,
    "args": {
      "External id": 2936,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1631, "finished": true, "Ev Idx": 2935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564838542.123, "dur": 11.402,
    "args": {
      "External id": 2937,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1631, "finished": true, "Ev Idx": 2936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564838544.381, "dur": 1.745,
    "args": {
      "External id": 2938,"Record function id": 0, "finished": true, "Ev Idx": 2937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564838544.720, "dur": 1.272,
    "args": {
      "External id": 2939,"Record function id": 0, "finished": true, "Ev Idx": 2938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564838546.628, "dur": 6.567,
    "args": {
      "External id": 2940,"Record function id": 0, "finished": true, "Ev Idx": 2939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564838553.918, "dur": 0.908,
    "args": {
      "External id": 2941,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1632, "finished": true, "Ev Idx": 2940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564838557.168, "dur": 1.657,
    "args": {
      "External id": 2942,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1633, "finished": true, "Ev Idx": 2941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838557.968, "dur": 0.395,
    "args": {
      "External id": 2943,"Record function id": 0, "finished": true, "Ev Idx": 2942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564838560.355, "dur": 47.265,
    "args": {
      "External id": 2944,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1634, "finished": true, "Ev Idx": 2943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564838561.085, "dur": 1.623,
    "args": {
      "External id": 2945,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1634, "finished": true, "Ev Idx": 2944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838561.941, "dur": 0.325,
    "args": {
      "External id": 2946,"Record function id": 0, "finished": true, "Ev Idx": 2945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838563.094, "dur": 4.183,
    "args": {
      "External id": 2947,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1635, "finished": true, "Ev Idx": 2946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564838565.165, "dur": 1.959,
    "args": {
      "External id": 2948,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1635, "finished": true, "Ev Idx": 2947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564838567.792, "dur": 2.452,
    "args": {
      "External id": 2949,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1636, "finished": true, "Ev Idx": 2948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838569.448, "dur": 0.424,
    "args": {
      "External id": 2950,"Record function id": 0, "finished": true, "Ev Idx": 2949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838570.560, "dur": 2.537,
    "args": {
      "External id": 2951,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1637, "finished": true, "Ev Idx": 2950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564838571.374, "dur": 1.615,
    "args": {
      "External id": 2952,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1637, "finished": true, "Ev Idx": 2951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564838573.848, "dur": 31.562,
    "args": {
      "External id": 2953,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1638, "finished": true, "Ev Idx": 2952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564838586.454, "dur": 1.012,
    "args": {
      "External id": 2954,"Record function id": 0, "finished": true, "Ev Idx": 2953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564838606.066, "dur": 1.121,
    "args": {
      "External id": 2955,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1639, "finished": true, "Ev Idx": 2954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564838611.262, "dur": 9.323,
    "args": {
      "External id": 2956,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1640, "finished": true, "Ev Idx": 2955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838623.664, "dur": 1.909,
    "args": {
      "External id": 2957,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1641, "finished": true, "Ev Idx": 2956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838624.497, "dur": 0.433,
    "args": {
      "External id": 2958,"Record function id": 0, "finished": true, "Ev Idx": 2957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838626.186, "dur": 1.290,
    "args": {
      "External id": 2959,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1641, "finished": true, "Ev Idx": 2958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838626.816, "dur": 0.306,
    "args": {
      "External id": 2960,"Record function id": 0, "finished": true, "Ev Idx": 2959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564838628.107, "dur": 1.066,
    "args": {
      "External id": 2961,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1641, "finished": true, "Ev Idx": 2960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838628.586, "dur": 0.248,
    "args": {
      "External id": 2962,"Record function id": 0, "finished": true, "Ev Idx": 2961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564838630.917, "dur": 7.466,
    "args": {
      "External id": 2963,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1641, "finished": true, "Ev Idx": 2962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564838643.991, "dur": 13.881,
    "args": {
      "External id": 2964,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1642, "finished": true, "Ev Idx": 2963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564838645.209, "dur": 0.391,
    "args": {
      "External id": 2965,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1642, "finished": true, "Ev Idx": 2964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564838646.359, "dur": 11.237,
    "args": {
      "External id": 2966,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1642, "finished": true, "Ev Idx": 2965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564838660.953, "dur": 0.199,
    "args": {
      "External id": 2967,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1643, "finished": true, "Ev Idx": 2966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564838673.123, "dur": 0.570,
    "args": {
      "External id": 2968,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1643, "finished": true, "Ev Idx": 2967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564838675.507, "dur": 40.186,
    "args": {
      "External id": 2969,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1643, "finished": true, "Ev Idx": 2968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564838675.952, "dur": 2.581,
    "args": {
      "External id": 2970,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1643, "finished": true, "Ev Idx": 2969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838677.635, "dur": 0.408,
    "args": {
      "External id": 2971,"Record function id": 0, "finished": true, "Ev Idx": 2970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838678.892, "dur": 2.629,
    "args": {
      "External id": 2972,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1644, "finished": true, "Ev Idx": 2971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838679.320, "dur": 2.094,
    "args": {
      "External id": 2973,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1644, "finished": true, "Ev Idx": 2972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564838681.936, "dur": 1.251,
    "args": {
      "External id": 2974,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1645, "finished": true, "Ev Idx": 2973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838682.592, "dur": 0.287,
    "args": {
      "External id": 2975,"Record function id": 0, "finished": true, "Ev Idx": 2974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838683.473, "dur": 1.376,
    "args": {
      "External id": 2976,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1646, "finished": true, "Ev Idx": 2975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838683.899, "dur": 0.881,
    "args": {
      "External id": 2977,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1646, "finished": true, "Ev Idx": 2976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564838685.232, "dur": 28.362,
    "args": {
      "External id": 2978,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1647, "finished": true, "Ev Idx": 2977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564838697.667, "dur": 0.980,
    "args": {
      "External id": 2979,"Record function id": 0, "finished": true, "Ev Idx": 2978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564838714.144, "dur": 1.186,
    "args": {
      "External id": 2980,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1648, "finished": true, "Ev Idx": 2979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564838717.490, "dur": 1.773,
    "args": {
      "External id": 2981,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1649, "finished": true, "Ev Idx": 2980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838718.295, "dur": 0.447,
    "args": {
      "External id": 2982,"Record function id": 0, "finished": true, "Ev Idx": 2981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564838722.060, "dur": 12.125,
    "args": {
      "External id": 2983,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1650, "finished": true, "Ev Idx": 2982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564838722.742, "dur": 11.302,
    "args": {
      "External id": 2984,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1650, "finished": true, "Ev Idx": 2983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564838723.829, "dur": 2.081,
    "args": {
      "External id": 2985,"Record function id": 0, "finished": true, "Ev Idx": 2984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564838724.250, "dur": 1.518,
    "args": {
      "External id": 2986,"Record function id": 0, "finished": true, "Ev Idx": 2985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564838726.249, "dur": 7.365,
    "args": {
      "External id": 2987,"Record function id": 0, "finished": true, "Ev Idx": 2986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838740.096, "dur": 1.771,
    "args": {
      "External id": 2988,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1651, "finished": true, "Ev Idx": 2987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838740.419, "dur": 1.340,
    "args": {
      "External id": 2989,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1651, "finished": true, "Ev Idx": 2988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564838754.109, "dur": 36.333,
    "args": {
      "External id": 2990,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1652, "finished": true, "Ev Idx": 2989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564838754.739, "dur": 2.912,
    "args": {
      "External id": 2991,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1652, "finished": true, "Ev Idx": 2990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564838756.004, "dur": 1.084,
    "args": {
      "External id": 2992,"Record function id": 0, "finished": true, "Ev Idx": 2991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838756.596, "dur": 0.390,
    "args": {
      "External id": 2993,"Record function id": 0, "finished": true, "Ev Idx": 2992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564838758.159, "dur": 32.134,
    "args": {
      "External id": 2994,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1653, "finished": true, "Ev Idx": 2993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838758.715, "dur": 2.296,
    "args": {
      "External id": 2995,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1653, "finished": true, "Ev Idx": 2994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838758.977, "dur": 1.956,
    "args": {
      "External id": 2996,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1653, "finished": true, "Ev Idx": 2995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564838761.419, "dur": 26.845,
    "args": {
      "External id": 2997,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1654, "finished": true, "Ev Idx": 2996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564838773.227, "dur": 1.142,
    "args": {
      "External id": 2998,"Record function id": 0, "finished": true, "Ev Idx": 2997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564838788.899, "dur": 1.108,
    "args": {
      "External id": 2999,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1655, "finished": true, "Ev Idx": 2998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564838795.246, "dur": 9.182,
    "args": {
      "External id": 3000,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1656, "finished": true, "Ev Idx": 2999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564838814.063, "dur": 0.346,
    "args": {
      "External id": 3001,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1657, "finished": true, "Ev Idx": 3000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564838816.475, "dur": 11.161,
    "args": {
      "External id": 3002,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1657, "finished": true, "Ev Idx": 3001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564838818.228, "dur": 0.408,
    "args": {
      "External id": 3003,"Record function id": 0, "finished": true, "Ev Idx": 3002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564838819.034, "dur": 0.063,
    "args": {
      "External id": 3004,"Record function id": 0, "finished": true, "Ev Idx": 3003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564838830.380, "dur": 14.541,
    "args": {
      "External id": 3005,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1658, "finished": true, "Ev Idx": 3004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564838848.587, "dur": 8.027,
    "args": {
      "External id": 3006,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1659, "finished": true, "Ev Idx": 3005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564838858.688, "dur": 11.068,
    "args": {
      "External id": 3007,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1660, "finished": true, "Ev Idx": 3006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564838871.621, "dur": 8.965,
    "args": {
      "External id": 3008,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1661, "finished": true, "Ev Idx": 3007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564838883.277, "dur": 0.100,
    "args": {
      "External id": 3009,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1662, "finished": true, "Ev Idx": 3008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564838884.429, "dur": 9.587,
    "args": {
      "External id": 3010,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1662, "finished": true, "Ev Idx": 3009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564838907.754, "dur": 105.034,
    "args": {
      "External id": 3011,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1663, "finished": true, "Ev Idx": 3010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564838908.431, "dur": 3.731,
    "args": {
      "External id": 3012,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1663, "finished": true, "Ev Idx": 3011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564838910.372, "dur": 1.088,
    "args": {
      "External id": 3013,"Record function id": 0, "finished": true, "Ev Idx": 3012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564838910.956, "dur": 0.384,
    "args": {
      "External id": 3014,"Record function id": 0, "finished": true, "Ev Idx": 3013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564838912.553, "dur": 100.102,
    "args": {
      "External id": 3015,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1664, "finished": true, "Ev Idx": 3014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564838912.977, "dur": 1.434,
    "args": {
      "External id": 3016,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1664, "finished": true, "Ev Idx": 3015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564838913.258, "dur": 1.050,
    "args": {
      "External id": 3017,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1664, "finished": true, "Ev Idx": 3016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564838914.709, "dur": 95.463,
    "args": {
      "External id": 3018,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1665, "finished": true, "Ev Idx": 3017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564838982.240, "dur": 1.264,
    "args": {
      "External id": 3019,"Record function id": 0, "finished": true, "Ev Idx": 3018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564839011.046, "dur": 1.176,
    "args": {
      "External id": 3020,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1666, "finished": true, "Ev Idx": 3019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564839021.955, "dur": 10.290,
    "args": {
      "External id": 3021,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1667, "finished": true, "Ev Idx": 3020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564839040.276, "dur": 75.027,
    "args": {
      "External id": 3022,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1668, "finished": true, "Ev Idx": 3021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564839040.724, "dur": 2.728,
    "args": {
      "External id": 3023,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1668, "finished": true, "Ev Idx": 3022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839041.824, "dur": 1.003,
    "args": {
      "External id": 3024,"Record function id": 0, "finished": true, "Ev Idx": 3023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839042.280, "dur": 0.427,
    "args": {
      "External id": 3025,"Record function id": 0, "finished": true, "Ev Idx": 3024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564839043.982, "dur": 71.213,
    "args": {
      "External id": 3026,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1669, "finished": true, "Ev Idx": 3025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839044.431, "dur": 2.578,
    "args": {
      "External id": 3027,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1669, "finished": true, "Ev Idx": 3026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839044.807, "dur": 2.108,
    "args": {
      "External id": 3028,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1669, "finished": true, "Ev Idx": 3027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564839047.346, "dur": 65.726,
    "args": {
      "External id": 3029,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1670, "finished": true, "Ev Idx": 3028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839089.689, "dur": 1.052,
    "args": {
      "External id": 3030,"Record function id": 0, "finished": true, "Ev Idx": 3029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564839113.654, "dur": 1.181,
    "args": {
      "External id": 3031,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1671, "finished": true, "Ev Idx": 3030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564839117.318, "dur": 8.955,
    "args": {
      "External id": 3032,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1672, "finished": true, "Ev Idx": 3031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564839133.318, "dur": 38.342,
    "args": {
      "External id": 3033,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1673, "finished": true, "Ev Idx": 3032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564839133.976, "dur": 2.452,
    "args": {
      "External id": 3034,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1673, "finished": true, "Ev Idx": 3033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839134.911, "dur": 0.897,
    "args": {
      "External id": 3035,"Record function id": 0, "finished": true, "Ev Idx": 3034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839135.325, "dur": 0.358,
    "args": {
      "External id": 3036,"Record function id": 0, "finished": true, "Ev Idx": 3035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564839136.712, "dur": 34.847,
    "args": {
      "External id": 3037,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1674, "finished": true, "Ev Idx": 3036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839137.155, "dur": 2.119,
    "args": {
      "External id": 3038,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1674, "finished": true, "Ev Idx": 3037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839137.456, "dur": 1.715,
    "args": {
      "External id": 3039,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1674, "finished": true, "Ev Idx": 3038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564839139.580, "dur": 30.252,
    "args": {
      "External id": 3040,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1675, "finished": true, "Ev Idx": 3039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839149.667, "dur": 1.015,
    "args": {
      "External id": 3041,"Record function id": 0, "finished": true, "Ev Idx": 3040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564839170.348, "dur": 0.900,
    "args": {
      "External id": 3042,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1676, "finished": true, "Ev Idx": 3041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564839174.628, "dur": 10.041,
    "args": {
      "External id": 3043,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1677, "finished": true, "Ev Idx": 3042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564839203.629, "dur": 0.181,
    "args": {
      "External id": 3044,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1678, "finished": true, "Ev Idx": 3043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564839205.585, "dur": 9.864,
    "args": {
      "External id": 3045,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1678, "finished": true, "Ev Idx": 3044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564839206.879, "dur": 0.429,
    "args": {
      "External id": 3046,"Record function id": 0, "finished": true, "Ev Idx": 3045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564839207.603, "dur": 0.071,
    "args": {
      "External id": 3047,"Record function id": 0, "finished": true, "Ev Idx": 3046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564839218.431, "dur": 12.070,
    "args": {
      "External id": 3048,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1679, "finished": true, "Ev Idx": 3047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564839234.250, "dur": 7.498,
    "args": {
      "External id": 3049,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1680, "finished": true, "Ev Idx": 3048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564839243.362, "dur": 8.104,
    "args": {
      "External id": 3050,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1681, "finished": true, "Ev Idx": 3049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564839253.106, "dur": 8.741,
    "args": {
      "External id": 3051,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1682, "finished": true, "Ev Idx": 3050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564839264.414, "dur": 0.282,
    "args": {
      "External id": 3052,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1683, "finished": true, "Ev Idx": 3051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564839265.546, "dur": 8.439,
    "args": {
      "External id": 3053,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1683, "finished": true, "Ev Idx": 3052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564839290.514, "dur": 37.730,
    "args": {
      "External id": 3054,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1684, "finished": true, "Ev Idx": 3053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564839291.038, "dur": 4.620,
    "args": {
      "External id": 3055,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1684, "finished": true, "Ev Idx": 3054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839293.446, "dur": 1.280,
    "args": {
      "External id": 3056,"Record function id": 0, "finished": true, "Ev Idx": 3055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839294.135, "dur": 0.471,
    "args": {
      "External id": 3057,"Record function id": 0, "finished": true, "Ev Idx": 3056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564839296.006, "dur": 32.137,
    "args": {
      "External id": 3058,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1685, "finished": true, "Ev Idx": 3057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839296.560, "dur": 1.753,
    "args": {
      "External id": 3059,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1685, "finished": true, "Ev Idx": 3058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839296.940, "dur": 1.253,
    "args": {
      "External id": 3060,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1685, "finished": true, "Ev Idx": 3059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564839298.670, "dur": 27.363,
    "args": {
      "External id": 3061,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1686, "finished": true, "Ev Idx": 3060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839309.904, "dur": 1.085,
    "args": {
      "External id": 3062,"Record function id": 0, "finished": true, "Ev Idx": 3061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564839326.711, "dur": 1.069,
    "args": {
      "External id": 3063,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1687, "finished": true, "Ev Idx": 3062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839331.413, "dur": 1.299,
    "args": {
      "External id": 3064,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1688, "finished": true, "Ev Idx": 3063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839334.622, "dur": 2.772,
    "args": {
      "External id": 3065,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1689, "finished": true, "Ev Idx": 3064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839335.977, "dur": 0.602,
    "args": {
      "External id": 3066,"Record function id": 0, "finished": true, "Ev Idx": 3065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564839344.999, "dur": 72.915,
    "args": {
      "External id": 3067,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1690, "finished": true, "Ev Idx": 3066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564839345.285, "dur": 2.854,
    "args": {
      "External id": 3068,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1690, "finished": true, "Ev Idx": 3067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839346.452, "dur": 1.246,
    "args": {
      "External id": 3069,"Record function id": 0, "finished": true, "Ev Idx": 3068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839347.097, "dur": 0.498,
    "args": {
      "External id": 3070,"Record function id": 0, "finished": true, "Ev Idx": 3069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564839348.684, "dur": 69.131,
    "args": {
      "External id": 3071,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1691, "finished": true, "Ev Idx": 3070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839349.214, "dur": 33.379,
    "args": {
      "External id": 3072,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1691, "finished": true, "Ev Idx": 3071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839381.505, "dur": 1.000,
    "args": {
      "External id": 3073,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1691, "finished": true, "Ev Idx": 3072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564839382.970, "dur": 33.015,
    "args": {
      "External id": 3074,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1692, "finished": true, "Ev Idx": 3073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839393.109, "dur": 1.082,
    "args": {
      "External id": 3075,"Record function id": 0, "finished": true, "Ev Idx": 3074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564839416.542, "dur": 0.976,
    "args": {
      "External id": 3076,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1693, "finished": true, "Ev Idx": 3075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839420.318, "dur": 1.149,
    "args": {
      "External id": 3077,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1694, "finished": true, "Ev Idx": 3076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839422.900, "dur": 1.657,
    "args": {
      "External id": 3078,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1695, "finished": true, "Ev Idx": 3077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839423.600, "dur": 0.450,
    "args": {
      "External id": 3079,"Record function id": 0, "finished": true, "Ev Idx": 3078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564839431.751, "dur": 36.146,
    "args": {
      "External id": 3080,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1696, "finished": true, "Ev Idx": 3079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564839432.042, "dur": 2.099,
    "args": {
      "External id": 3081,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1696, "finished": true, "Ev Idx": 3080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839432.960, "dur": 0.810,
    "args": {
      "External id": 3082,"Record function id": 0, "finished": true, "Ev Idx": 3081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839433.397, "dur": 0.266,
    "args": {
      "External id": 3083,"Record function id": 0, "finished": true, "Ev Idx": 3082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564839434.435, "dur": 33.367,
    "args": {
      "External id": 3084,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1697, "finished": true, "Ev Idx": 3083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839434.792, "dur": 1.168,
    "args": {
      "External id": 3085,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1697, "finished": true, "Ev Idx": 3084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839435.055, "dur": 0.817,
    "args": {
      "External id": 3086,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1697, "finished": true, "Ev Idx": 3085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564839436.264, "dur": 29.882,
    "args": {
      "External id": 3087,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1698, "finished": true, "Ev Idx": 3086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839446.472, "dur": 1.039,
    "args": {
      "External id": 3088,"Record function id": 0, "finished": true, "Ev Idx": 3087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564839466.626, "dur": 0.905,
    "args": {
      "External id": 3089,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1699, "finished": true, "Ev Idx": 3088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839470.175, "dur": 1.116,
    "args": {
      "External id": 3090,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1700, "finished": true, "Ev Idx": 3089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839472.891, "dur": 1.665,
    "args": {
      "External id": 3091,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1701, "finished": true, "Ev Idx": 3090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839473.654, "dur": 0.405,
    "args": {
      "External id": 3092,"Record function id": 0, "finished": true, "Ev Idx": 3091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564839477.539, "dur": 2.105,
    "args": {
      "External id": 3093,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1702, "finished": true, "Ev Idx": 3092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839478.836, "dur": 0.278,
    "args": {
      "External id": 3094,"Record function id": 0, "finished": true, "Ev Idx": 3093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564839480.835, "dur": 1.125,
    "args": {
      "External id": 3095,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1702, "finished": true, "Ev Idx": 3094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839481.244, "dur": 0.268,
    "args": {
      "External id": 3096,"Record function id": 0, "finished": true, "Ev Idx": 3095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564839483.002, "dur": 10.071,
    "args": {
      "External id": 3097,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1702, "finished": true, "Ev Idx": 3096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839498.496, "dur": 4.162,
    "args": {
      "External id": 3098,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1703, "finished": true, "Ev Idx": 3097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839500.720, "dur": 0.500,
    "args": {
      "External id": 3099,"Record function id": 0, "finished": true, "Ev Idx": 3098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839504.918, "dur": 3.245,
    "args": {
      "External id": 3100,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1704, "finished": true, "Ev Idx": 3099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839506.603, "dur": 0.607,
    "args": {
      "External id": 3101,"Record function id": 0, "finished": true, "Ev Idx": 3100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564839510.016, "dur": 9.792,
    "args": {
      "External id": 3102,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1705, "finished": true, "Ev Idx": 3101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564839524.080, "dur": 17.703,
    "args": {
      "External id": 3103,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1706, "finished": true, "Ev Idx": 3102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564839544.681, "dur": 54.905,
    "args": {
      "External id": 3104,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1707, "finished": true, "Ev Idx": 3103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564839602.323, "dur": 10.103,
    "args": {
      "External id": 3105,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1708, "finished": true, "Ev Idx": 3104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564839614.604, "dur": 7.790,
    "args": {
      "External id": 3106,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1709, "finished": true, "Ev Idx": 3105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839625.271, "dur": 2.522,
    "args": {
      "External id": 3107,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1710, "finished": true, "Ev Idx": 3106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839626.421, "dur": 0.528,
    "args": {
      "External id": 3108,"Record function id": 0, "finished": true, "Ev Idx": 3107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839629.411, "dur": 1.699,
    "args": {
      "External id": 3109,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1711, "finished": true, "Ev Idx": 3108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839630.228, "dur": 0.335,
    "args": {
      "External id": 3110,"Record function id": 0, "finished": true, "Ev Idx": 3109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564839632.473, "dur": 8.398,
    "args": {
      "External id": 3111,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1712, "finished": true, "Ev Idx": 3110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564839643.158, "dur": 10.428,
    "args": {
      "External id": 3112,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1713, "finished": true, "Ev Idx": 3111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564839655.927, "dur": 8.819,
    "args": {
      "External id": 3113,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1714, "finished": true, "Ev Idx": 3112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564839666.060, "dur": 7.276,
    "args": {
      "External id": 3114,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1715, "finished": true, "Ev Idx": 3113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839698.240, "dur": 2.623,
    "args": {
      "External id": 3115,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1716, "finished": true, "Ev Idx": 3114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839699.634, "dur": 0.472,
    "args": {
      "External id": 3116,"Record function id": 0, "finished": true, "Ev Idx": 3115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839701.531, "dur": 1.527,
    "args": {
      "External id": 3117,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1717, "finished": true, "Ev Idx": 3116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839702.259, "dur": 0.334,
    "args": {
      "External id": 3118,"Record function id": 0, "finished": true, "Ev Idx": 3117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564839704.030, "dur": 3.586,
    "args": {
      "External id": 3119,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1718, "finished": true, "Ev Idx": 3118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839706.102, "dur": 0.818,
    "args": {
      "External id": 3120,"Record function id": 0, "finished": true, "Ev Idx": 3119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839708.464, "dur": 3.129,
    "args": {
      "External id": 3121,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1719, "finished": true, "Ev Idx": 3120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839710.598, "dur": 0.483,
    "args": {
      "External id": 3122,"Record function id": 0, "finished": true, "Ev Idx": 3121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839712.221, "dur": 2.116,
    "args": {
      "External id": 3123,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1720, "finished": true, "Ev Idx": 3122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839713.603, "dur": 0.281,
    "args": {
      "External id": 3124,"Record function id": 0, "finished": true, "Ev Idx": 3123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564839717.312, "dur": 3.338,
    "args": {
      "External id": 3125,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1721, "finished": true, "Ev Idx": 3124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839719.384, "dur": 0.555,
    "args": {
      "External id": 3126,"Record function id": 0, "finished": true, "Ev Idx": 3125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839723.124, "dur": 20.060,
    "args": {
      "External id": 3127,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1722, "finished": true, "Ev Idx": 3126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564839725.913, "dur": 15.022,
    "args": {
      "External id": 3128,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1722, "finished": true, "Ev Idx": 3127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564839727.631, "dur": 2.904,
    "args": {
      "External id": 3129,"Record function id": 0, "finished": true, "Ev Idx": 3128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839728.565, "dur": 1.705,
    "args": {
      "External id": 3130,"Record function id": 0, "finished": true, "Ev Idx": 3129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564839731.216, "dur": 9.128,
    "args": {
      "External id": 3131,"Record function id": 0, "finished": true, "Ev Idx": 3130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564839741.444, "dur": 1.480,
    "args": {
      "External id": 3132,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1723, "finished": true, "Ev Idx": 3131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839746.391, "dur": 2.458,
    "args": {
      "External id": 3133,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1724, "finished": true, "Ev Idx": 3132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839747.873, "dur": 0.376,
    "args": {
      "External id": 3134,"Record function id": 0, "finished": true, "Ev Idx": 3133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839749.473, "dur": 2.527,
    "args": {
      "External id": 3135,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1725, "finished": true, "Ev Idx": 3134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839751.141, "dur": 0.407,
    "args": {
      "External id": 3136,"Record function id": 0, "finished": true, "Ev Idx": 3135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564839752.534, "dur": 1.558,
    "args": {
      "External id": 3137,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1726, "finished": true, "Ev Idx": 3136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839753.322, "dur": 0.355,
    "args": {
      "External id": 3138,"Record function id": 0, "finished": true, "Ev Idx": 3137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839754.663, "dur": 1.438,
    "args": {
      "External id": 3139,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1727, "finished": true, "Ev Idx": 3138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839755.353, "dur": 0.297,
    "args": {
      "External id": 3140,"Record function id": 0, "finished": true, "Ev Idx": 3139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839756.671, "dur": 1.554,
    "args": {
      "External id": 3141,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1728, "finished": true, "Ev Idx": 3140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839757.511, "dur": 0.277,
    "args": {
      "External id": 3142,"Record function id": 0, "finished": true, "Ev Idx": 3141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564839759.928, "dur": 1.762,
    "args": {
      "External id": 3143,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1729, "finished": true, "Ev Idx": 3142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839760.937, "dur": 0.336,
    "args": {
      "External id": 3144,"Record function id": 0, "finished": true, "Ev Idx": 3143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839763.115, "dur": 13.395,
    "args": {
      "External id": 3145,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1730, "finished": true, "Ev Idx": 3144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564839764.179, "dur": 10.180,
    "args": {
      "External id": 3146,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1730, "finished": true, "Ev Idx": 3145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564839765.124, "dur": 1.929,
    "args": {
      "External id": 3147,"Record function id": 0, "finished": true, "Ev Idx": 3146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839765.646, "dur": 1.275,
    "args": {
      "External id": 3148,"Record function id": 0, "finished": true, "Ev Idx": 3147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564839767.383, "dur": 6.590,
    "args": {
      "External id": 3149,"Record function id": 0, "finished": true, "Ev Idx": 3148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564839774.778, "dur": 1.541,
    "args": {
      "External id": 3150,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1731, "finished": true, "Ev Idx": 3149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839778.778, "dur": 1.850,
    "args": {
      "External id": 3151,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1732, "finished": true, "Ev Idx": 3150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839779.684, "dur": 0.403,
    "args": {
      "External id": 3152,"Record function id": 0, "finished": true, "Ev Idx": 3151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564839782.143, "dur": 48.057,
    "args": {
      "External id": 3153,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1733, "finished": true, "Ev Idx": 3152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564839782.718, "dur": 1.655,
    "args": {
      "External id": 3154,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1733, "finished": true, "Ev Idx": 3153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839783.589, "dur": 0.332,
    "args": {
      "External id": 3155,"Record function id": 0, "finished": true, "Ev Idx": 3154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839784.714, "dur": 3.942,
    "args": {
      "External id": 3156,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1734, "finished": true, "Ev Idx": 3155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564839786.627, "dur": 1.854,
    "args": {
      "External id": 3157,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1734, "finished": true, "Ev Idx": 3156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564839789.186, "dur": 1.332,
    "args": {
      "External id": 3158,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1735, "finished": true, "Ev Idx": 3157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839789.888, "dur": 0.272,
    "args": {
      "External id": 3159,"Record function id": 0, "finished": true, "Ev Idx": 3158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839790.852, "dur": 3.188,
    "args": {
      "External id": 3160,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1736, "finished": true, "Ev Idx": 3159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564839791.732, "dur": 2.209,
    "args": {
      "External id": 3161,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1736, "finished": true, "Ev Idx": 3160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564839794.734, "dur": 33.314,
    "args": {
      "External id": 3162,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1737, "finished": true, "Ev Idx": 3161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839809.317, "dur": 1.199,
    "args": {
      "External id": 3163,"Record function id": 0, "finished": true, "Ev Idx": 3162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564839828.618, "dur": 1.194,
    "args": {
      "External id": 3164,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1738, "finished": true, "Ev Idx": 3163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564839833.882, "dur": 9.667,
    "args": {
      "External id": 3165,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1739, "finished": true, "Ev Idx": 3164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839846.525, "dur": 1.977,
    "args": {
      "External id": 3166,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1740, "finished": true, "Ev Idx": 3165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839847.263, "dur": 0.532,
    "args": {
      "External id": 3167,"Record function id": 0, "finished": true, "Ev Idx": 3166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839849.167, "dur": 1.304,
    "args": {
      "External id": 3168,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1740, "finished": true, "Ev Idx": 3167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839849.777, "dur": 0.279,
    "args": {
      "External id": 3169,"Record function id": 0, "finished": true, "Ev Idx": 3168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564839851.111, "dur": 1.023,
    "args": {
      "External id": 3170,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1740, "finished": true, "Ev Idx": 3169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839851.577, "dur": 0.236,
    "args": {
      "External id": 3171,"Record function id": 0, "finished": true, "Ev Idx": 3170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564839853.871, "dur": 8.575,
    "args": {
      "External id": 3172,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1740, "finished": true, "Ev Idx": 3171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564839867.808, "dur": 13.285,
    "args": {
      "External id": 3173,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1741, "finished": true, "Ev Idx": 3172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564839868.742, "dur": 0.394,
    "args": {
      "External id": 3174,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1741, "finished": true, "Ev Idx": 3173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564839869.780, "dur": 11.037,
    "args": {
      "External id": 3175,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1741, "finished": true, "Ev Idx": 3174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564839884.305, "dur": 0.219,
    "args": {
      "External id": 3176,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1742, "finished": true, "Ev Idx": 3175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564839896.138, "dur": 0.440,
    "args": {
      "External id": 3177,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1742, "finished": true, "Ev Idx": 3176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564839898.495, "dur": 41.357,
    "args": {
      "External id": 3178,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1742, "finished": true, "Ev Idx": 3177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564839899.172, "dur": 2.668,
    "args": {
      "External id": 3179,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1742, "finished": true, "Ev Idx": 3178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839900.859, "dur": 0.368,
    "args": {
      "External id": 3180,"Record function id": 0, "finished": true, "Ev Idx": 3179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839902.234, "dur": 2.332,
    "args": {
      "External id": 3181,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1743, "finished": true, "Ev Idx": 3180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839902.665, "dur": 1.786,
    "args": {
      "External id": 3182,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1743, "finished": true, "Ev Idx": 3181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564839904.973, "dur": 1.218,
    "args": {
      "External id": 3183,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1744, "finished": true, "Ev Idx": 3182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839905.574, "dur": 0.297,
    "args": {
      "External id": 3184,"Record function id": 0, "finished": true, "Ev Idx": 3183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839906.477, "dur": 1.172,
    "args": {
      "External id": 3185,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1745, "finished": true, "Ev Idx": 3184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839906.748, "dur": 0.831,
    "args": {
      "External id": 3186,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1745, "finished": true, "Ev Idx": 3185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564839908.062, "dur": 29.782,
    "args": {
      "External id": 3187,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1746, "finished": true, "Ev Idx": 3186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839920.054, "dur": 1.097,
    "args": {
      "External id": 3188,"Record function id": 0, "finished": true, "Ev Idx": 3187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564839938.473, "dur": 1.057,
    "args": {
      "External id": 3189,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1747, "finished": true, "Ev Idx": 3188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839942.016, "dur": 1.748,
    "args": {
      "External id": 3190,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1748, "finished": true, "Ev Idx": 3189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839942.830, "dur": 0.372,
    "args": {
      "External id": 3191,"Record function id": 0, "finished": true, "Ev Idx": 3190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564839946.314, "dur": 11.110,
    "args": {
      "External id": 3192,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1749, "finished": true, "Ev Idx": 3191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564839946.869, "dur": 10.411,
    "args": {
      "External id": 3193,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1749, "finished": true, "Ev Idx": 3192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564839948.056, "dur": 1.943,
    "args": {
      "External id": 3194,"Record function id": 0, "finished": true, "Ev Idx": 3193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839948.420, "dur": 1.378,
    "args": {
      "External id": 3195,"Record function id": 0, "finished": true, "Ev Idx": 3194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564839950.342, "dur": 6.534,
    "args": {
      "External id": 3196,"Record function id": 0, "finished": true, "Ev Idx": 3195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839963.249, "dur": 1.854,
    "args": {
      "External id": 3197,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1750, "finished": true, "Ev Idx": 3196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839963.686, "dur": 1.312,
    "args": {
      "External id": 3198,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1750, "finished": true, "Ev Idx": 3197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564839979.491, "dur": 38.449,
    "args": {
      "External id": 3199,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1751, "finished": true, "Ev Idx": 3198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564839980.238, "dur": 3.449,
    "args": {
      "External id": 3200,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1751, "finished": true, "Ev Idx": 3199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564839981.862, "dur": 1.166,
    "args": {
      "External id": 3201,"Record function id": 0, "finished": true, "Ev Idx": 3200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564839982.385, "dur": 0.516,
    "args": {
      "External id": 3202,"Record function id": 0, "finished": true, "Ev Idx": 3201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564839984.038, "dur": 33.729,
    "args": {
      "External id": 3203,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1752, "finished": true, "Ev Idx": 3202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564839984.689, "dur": 1.547,
    "args": {
      "External id": 3204,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1752, "finished": true, "Ev Idx": 3203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564839984.953, "dur": 1.187,
    "args": {
      "External id": 3205,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1752, "finished": true, "Ev Idx": 3204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564839986.697, "dur": 29.099,
    "args": {
      "External id": 3206,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1753, "finished": true, "Ev Idx": 3205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564839999.419, "dur": 1.140,
    "args": {
      "External id": 3207,"Record function id": 0, "finished": true, "Ev Idx": 3206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564840016.374, "dur": 1.046,
    "args": {
      "External id": 3208,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1754, "finished": true, "Ev Idx": 3207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564840022.919, "dur": 8.883,
    "args": {
      "External id": 3209,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1755, "finished": true, "Ev Idx": 3208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564840040.954, "dur": 0.199,
    "args": {
      "External id": 3210,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1756, "finished": true, "Ev Idx": 3209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564840043.444, "dur": 13.106,
    "args": {
      "External id": 3211,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1756, "finished": true, "Ev Idx": 3210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564840045.477, "dur": 0.364,
    "args": {
      "External id": 3212,"Record function id": 0, "finished": true, "Ev Idx": 3211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564840046.317, "dur": 0.080,
    "args": {
      "External id": 3213,"Record function id": 0, "finished": true, "Ev Idx": 3212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564840059.624, "dur": 12.827,
    "args": {
      "External id": 3214,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1757, "finished": true, "Ev Idx": 3213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564840076.213, "dur": 9.013,
    "args": {
      "External id": 3215,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1758, "finished": true, "Ev Idx": 3214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564840087.302, "dur": 8.726,
    "args": {
      "External id": 3216,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1759, "finished": true, "Ev Idx": 3215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564840097.783, "dur": 9.508,
    "args": {
      "External id": 3217,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1760, "finished": true, "Ev Idx": 3216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564840110.293, "dur": 0.119,
    "args": {
      "External id": 3218,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1761, "finished": true, "Ev Idx": 3217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564840111.686, "dur": 8.317,
    "args": {
      "External id": 3219,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1761, "finished": true, "Ev Idx": 3218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564840136.504, "dur": 47.332,
    "args": {
      "External id": 3220,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1762, "finished": true, "Ev Idx": 3219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564840137.123, "dur": 4.250,
    "args": {
      "External id": 3221,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1762, "finished": true, "Ev Idx": 3220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564840139.358, "dur": 1.416,
    "args": {
      "External id": 3222,"Record function id": 0, "finished": true, "Ev Idx": 3221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840140.190, "dur": 0.459,
    "args": {
      "External id": 3223,"Record function id": 0, "finished": true, "Ev Idx": 3222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564840141.887, "dur": 41.816,
    "args": {
      "External id": 3224,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1763, "finished": true, "Ev Idx": 3223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564840142.349, "dur": 2.180,
    "args": {
      "External id": 3225,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1763, "finished": true, "Ev Idx": 3224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564840142.830, "dur": 1.604,
    "args": {
      "External id": 3226,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1763, "finished": true, "Ev Idx": 3225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564840144.871, "dur": 36.824,
    "args": {
      "External id": 3227,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1764, "finished": true, "Ev Idx": 3226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564840154.348, "dur": 1.273,
    "args": {
      "External id": 3228,"Record function id": 0, "finished": true, "Ev Idx": 3227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564840182.350, "dur": 1.035,
    "args": {
      "External id": 3229,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1765, "finished": true, "Ev Idx": 3228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564840192.677, "dur": 59.434,
    "args": {
      "External id": 3230,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1766, "finished": true, "Ev Idx": 3229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564840260.969, "dur": 45.507,
    "args": {
      "External id": 3231,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1767, "finished": true, "Ev Idx": 3230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564840261.526, "dur": 3.520,
    "args": {
      "External id": 3232,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1767, "finished": true, "Ev Idx": 3231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564840263.014, "dur": 1.380,
    "args": {
      "External id": 3233,"Record function id": 0, "finished": true, "Ev Idx": 3232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840263.661, "dur": 0.606,
    "args": {
      "External id": 3234,"Record function id": 0, "finished": true, "Ev Idx": 3233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564840265.386, "dur": 40.981,
    "args": {
      "External id": 3235,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1768, "finished": true, "Ev Idx": 3234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564840265.852, "dur": 2.108,
    "args": {
      "External id": 3236,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1768, "finished": true, "Ev Idx": 3235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564840266.166, "dur": 1.677,
    "args": {
      "External id": 3237,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1768, "finished": true, "Ev Idx": 3236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564840268.328, "dur": 36.130,
    "args": {
      "External id": 3238,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1769, "finished": true, "Ev Idx": 3237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564840280.339, "dur": 1.183,
    "args": {
      "External id": 3239,"Record function id": 0, "finished": true, "Ev Idx": 3238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564840305.060, "dur": 0.978,
    "args": {
      "External id": 3240,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1770, "finished": true, "Ev Idx": 3239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564840308.477, "dur": 41.235,
    "args": {
      "External id": 3241,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1771, "finished": true, "Ev Idx": 3240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564840357.792, "dur": 44.345,
    "args": {
      "External id": 3242,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1772, "finished": true, "Ev Idx": 3241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564840358.182, "dur": 4.029,
    "args": {
      "External id": 3243,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1772, "finished": true, "Ev Idx": 3242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564840360.495, "dur": 1.126,
    "args": {
      "External id": 3244,"Record function id": 0, "finished": true, "Ev Idx": 3243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840361.042, "dur": 0.447,
    "args": {
      "External id": 3245,"Record function id": 0, "finished": true, "Ev Idx": 3244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564840362.532, "dur": 39.515,
    "args": {
      "External id": 3246,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1773, "finished": true, "Ev Idx": 3245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564840362.958, "dur": 1.531,
    "args": {
      "External id": 3247,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1773, "finished": true, "Ev Idx": 3246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564840363.267, "dur": 1.107,
    "args": {
      "External id": 3248,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1773, "finished": true, "Ev Idx": 3247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564840364.848, "dur": 35.316,
    "args": {
      "External id": 3249,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1774, "finished": true, "Ev Idx": 3248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564840376.708, "dur": 1.162,
    "args": {
      "External id": 3250,"Record function id": 0, "finished": true, "Ev Idx": 3249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564840400.763, "dur": 0.960,
    "args": {
      "External id": 3251,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1775, "finished": true, "Ev Idx": 3250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564840404.756, "dur": 9.424,
    "args": {
      "External id": 3252,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1776, "finished": true, "Ev Idx": 3251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564840434.713, "dur": 0.214,
    "args": {
      "External id": 3253,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1777, "finished": true, "Ev Idx": 3252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564840436.873, "dur": 9.857,
    "args": {
      "External id": 3254,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1777, "finished": true, "Ev Idx": 3253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564840438.282, "dur": 0.394,
    "args": {
      "External id": 3255,"Record function id": 0, "finished": true, "Ev Idx": 3254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564840438.998, "dur": 0.067,
    "args": {
      "External id": 3256,"Record function id": 0, "finished": true, "Ev Idx": 3255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564840449.364, "dur": 14.815,
    "args": {
      "External id": 3257,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1778, "finished": true, "Ev Idx": 3256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564840468.033, "dur": 8.199,
    "args": {
      "External id": 3258,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1779, "finished": true, "Ev Idx": 3257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564840478.036, "dur": 11.283,
    "args": {
      "External id": 3259,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1780, "finished": true, "Ev Idx": 3258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564840490.949, "dur": 8.284,
    "args": {
      "External id": 3260,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1781, "finished": true, "Ev Idx": 3259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564840501.584, "dur": 0.119,
    "args": {
      "External id": 3261,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1782, "finished": true, "Ev Idx": 3260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564840502.574, "dur": 8.610,
    "args": {
      "External id": 3262,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1782, "finished": true, "Ev Idx": 3261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564840528.662, "dur": 38.105,
    "args": {
      "External id": 3263,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1783, "finished": true, "Ev Idx": 3262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564840529.130, "dur": 4.873,
    "args": {
      "External id": 3264,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1783, "finished": true, "Ev Idx": 3263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564840531.309, "dur": 1.638,
    "args": {
      "External id": 3265,"Record function id": 0, "finished": true, "Ev Idx": 3264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840532.268, "dur": 0.548,
    "args": {
      "External id": 3266,"Record function id": 0, "finished": true, "Ev Idx": 3265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564840534.315, "dur": 32.354,
    "args": {
      "External id": 3267,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1784, "finished": true, "Ev Idx": 3266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564840534.958, "dur": 1.952,
    "args": {
      "External id": 3268,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1784, "finished": true, "Ev Idx": 3267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564840535.279, "dur": 1.511,
    "args": {
      "External id": 3269,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1784, "finished": true, "Ev Idx": 3268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564840537.563, "dur": 27.041,
    "args": {
      "External id": 3270,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1785, "finished": true, "Ev Idx": 3269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564840549.293, "dur": 1.042,
    "args": {
      "External id": 3271,"Record function id": 0, "finished": true, "Ev Idx": 3270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564840565.186, "dur": 1.211,
    "args": {
      "External id": 3272,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1786, "finished": true, "Ev Idx": 3271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564840569.631, "dur": 1.405,
    "args": {
      "External id": 3273,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1787, "finished": true, "Ev Idx": 3272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564840572.704, "dur": 2.336,
    "args": {
      "External id": 3274,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1788, "finished": true, "Ev Idx": 3273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840573.790, "dur": 0.489,
    "args": {
      "External id": 3275,"Record function id": 0, "finished": true, "Ev Idx": 3274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564840581.463, "dur": 44.112,
    "args": {
      "External id": 3276,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1789, "finished": true, "Ev Idx": 3275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564840581.958, "dur": 2.678,
    "args": {
      "External id": 3277,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1789, "finished": true, "Ev Idx": 3276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564840583.140, "dur": 1.024,
    "args": {
      "External id": 3278,"Record function id": 0, "finished": true, "Ev Idx": 3277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840583.801, "dur": 0.273,
    "args": {
      "External id": 3279,"Record function id": 0, "finished": true, "Ev Idx": 3278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564840585.139, "dur": 40.336,
    "args": {
      "External id": 3280,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1790, "finished": true, "Ev Idx": 3279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564840585.517, "dur": 1.777,
    "args": {
      "External id": 3281,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1790, "finished": true, "Ev Idx": 3280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564840586.336, "dur": 0.861,
    "args": {
      "External id": 3282,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1790, "finished": true, "Ev Idx": 3281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564840587.590, "dur": 36.065,
    "args": {
      "External id": 3283,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1791, "finished": true, "Ev Idx": 3282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564840600.979, "dur": 1.243,
    "args": {
      "External id": 3284,"Record function id": 0, "finished": true, "Ev Idx": 3283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564840624.217, "dur": 0.978,
    "args": {
      "External id": 3285,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1792, "finished": true, "Ev Idx": 3284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564840627.731, "dur": 1.311,
    "args": {
      "External id": 3286,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1793, "finished": true, "Ev Idx": 3285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564840630.424, "dur": 1.624,
    "args": {
      "External id": 3287,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1794, "finished": true, "Ev Idx": 3286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840631.158, "dur": 0.409,
    "args": {
      "External id": 3288,"Record function id": 0, "finished": true, "Ev Idx": 3287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564840638.523, "dur": 38.458,
    "args": {
      "External id": 3289,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1795, "finished": true, "Ev Idx": 3288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564840639.000, "dur": 2.056,
    "args": {
      "External id": 3290,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1795, "finished": true, "Ev Idx": 3289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564840639.871, "dur": 0.779,
    "args": {
      "External id": 3291,"Record function id": 0, "finished": true, "Ev Idx": 3290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840640.271, "dur": 0.284,
    "args": {
      "External id": 3292,"Record function id": 0, "finished": true, "Ev Idx": 3291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564840641.334, "dur": 35.547,
    "args": {
      "External id": 3293,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1796, "finished": true, "Ev Idx": 3292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564840641.872, "dur": 2.680,
    "args": {
      "External id": 3294,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1796, "finished": true, "Ev Idx": 3293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564840642.141, "dur": 2.313,
    "args": {
      "External id": 3295,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1796, "finished": true, "Ev Idx": 3294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564840644.843, "dur": 30.187,
    "args": {
      "External id": 3296,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1797, "finished": true, "Ev Idx": 3295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564840654.163, "dur": 1.190,
    "args": {
      "External id": 3297,"Record function id": 0, "finished": true, "Ev Idx": 3296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564840675.738, "dur": 0.880,
    "args": {
      "External id": 3298,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1798, "finished": true, "Ev Idx": 3297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564840679.233, "dur": 1.117,
    "args": {
      "External id": 3299,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1799, "finished": true, "Ev Idx": 3298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564840681.768, "dur": 1.699,
    "args": {
      "External id": 3300,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1800, "finished": true, "Ev Idx": 3299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840682.489, "dur": 0.404,
    "args": {
      "External id": 3301,"Record function id": 0, "finished": true, "Ev Idx": 3300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564840686.136, "dur": 2.416,
    "args": {
      "External id": 3302,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1801, "finished": true, "Ev Idx": 3301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840687.606, "dur": 0.418,
    "args": {
      "External id": 3303,"Record function id": 0, "finished": true, "Ev Idx": 3302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564840689.555, "dur": 1.231,
    "args": {
      "External id": 3304,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1801, "finished": true, "Ev Idx": 3303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840690.170, "dur": 0.281,
    "args": {
      "External id": 3305,"Record function id": 0, "finished": true, "Ev Idx": 3304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564840691.858, "dur": 10.830,
    "args": {
      "External id": 3306,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1801, "finished": true, "Ev Idx": 3305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840708.324, "dur": 5.300,
    "args": {
      "External id": 3307,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1802, "finished": true, "Ev Idx": 3306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840711.371, "dur": 0.927,
    "args": {
      "External id": 3308,"Record function id": 0, "finished": true, "Ev Idx": 3307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840716.114, "dur": 2.094,
    "args": {
      "External id": 3309,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1803, "finished": true, "Ev Idx": 3308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840717.138, "dur": 0.486,
    "args": {
      "External id": 3310,"Record function id": 0, "finished": true, "Ev Idx": 3309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564840719.961, "dur": 12.759,
    "args": {
      "External id": 3311,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1804, "finished": true, "Ev Idx": 3310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564840736.820, "dur": 16.343,
    "args": {
      "External id": 3312,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1805, "finished": true, "Ev Idx": 3311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564840756.037, "dur": 9.145,
    "args": {
      "External id": 3313,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1806, "finished": true, "Ev Idx": 3312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564840767.052, "dur": 9.311,
    "args": {
      "External id": 3314,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1807, "finished": true, "Ev Idx": 3313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564840778.464, "dur": 6.657,
    "args": {
      "External id": 3315,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1808, "finished": true, "Ev Idx": 3314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840787.755, "dur": 3.126,
    "args": {
      "External id": 3316,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1809, "finished": true, "Ev Idx": 3315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840789.756, "dur": 0.437,
    "args": {
      "External id": 3317,"Record function id": 0, "finished": true, "Ev Idx": 3316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840792.449, "dur": 3.054,
    "args": {
      "External id": 3318,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1810, "finished": true, "Ev Idx": 3317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840794.480, "dur": 0.475,
    "args": {
      "External id": 3319,"Record function id": 0, "finished": true, "Ev Idx": 3318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564840796.525, "dur": 6.836,
    "args": {
      "External id": 3320,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1811, "finished": true, "Ev Idx": 3319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564840805.408, "dur": 12.173,
    "args": {
      "External id": 3321,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1812, "finished": true, "Ev Idx": 3320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564840819.827, "dur": 7.167,
    "args": {
      "External id": 3322,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1813, "finished": true, "Ev Idx": 3321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564840828.345, "dur": 6.211,
    "args": {
      "External id": 3323,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1814, "finished": true, "Ev Idx": 3322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840857.360, "dur": 3.078,
    "args": {
      "External id": 3324,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1815, "finished": true, "Ev Idx": 3323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840859.261, "dur": 0.476,
    "args": {
      "External id": 3325,"Record function id": 0, "finished": true, "Ev Idx": 3324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840861.099, "dur": 2.380,
    "args": {
      "External id": 3326,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1816, "finished": true, "Ev Idx": 3325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840862.671, "dur": 0.300,
    "args": {
      "External id": 3327,"Record function id": 0, "finished": true, "Ev Idx": 3326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564840864.165, "dur": 2.197,
    "args": {
      "External id": 3328,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1817, "finished": true, "Ev Idx": 3327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840865.119, "dur": 0.592,
    "args": {
      "External id": 3329,"Record function id": 0, "finished": true, "Ev Idx": 3328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840867.067, "dur": 1.784,
    "args": {
      "External id": 3330,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1818, "finished": true, "Ev Idx": 3329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840867.967, "dur": 0.362,
    "args": {
      "External id": 3331,"Record function id": 0, "finished": true, "Ev Idx": 3330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840869.489, "dur": 1.754,
    "args": {
      "External id": 3332,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1819, "finished": true, "Ev Idx": 3331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840870.337, "dur": 0.433,
    "args": {
      "External id": 3333,"Record function id": 0, "finished": true, "Ev Idx": 3332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564840873.986, "dur": 3.167,
    "args": {
      "External id": 3334,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1820, "finished": true, "Ev Idx": 3333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840875.874, "dur": 0.573,
    "args": {
      "External id": 3335,"Record function id": 0, "finished": true, "Ev Idx": 3334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564840879.774, "dur": 20.165,
    "args": {
      "External id": 3336,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1821, "finished": true, "Ev Idx": 3335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564840882.440, "dur": 15.072,
    "args": {
      "External id": 3337,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1821, "finished": true, "Ev Idx": 3336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564840883.823, "dur": 2.643,
    "args": {
      "External id": 3338,"Record function id": 0, "finished": true, "Ev Idx": 3337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564840884.666, "dur": 1.610,
    "args": {
      "External id": 3339,"Record function id": 0, "finished": true, "Ev Idx": 3338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564840887.124, "dur": 9.758,
    "args": {
      "External id": 3340,"Record function id": 0, "finished": true, "Ev Idx": 3339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564840898.003, "dur": 1.701,
    "args": {
      "External id": 3341,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1822, "finished": true, "Ev Idx": 3340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840903.296, "dur": 2.616,
    "args": {
      "External id": 3342,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1823, "finished": true, "Ev Idx": 3341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840904.979, "dur": 0.375,
    "args": {
      "External id": 3343,"Record function id": 0, "finished": true, "Ev Idx": 3342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840906.504, "dur": 1.470,
    "args": {
      "External id": 3344,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1824, "finished": true, "Ev Idx": 3343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840907.241, "dur": 0.294,
    "args": {
      "External id": 3345,"Record function id": 0, "finished": true, "Ev Idx": 3344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564840908.832, "dur": 1.780,
    "args": {
      "External id": 3346,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1825, "finished": true, "Ev Idx": 3345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840909.883, "dur": 0.315,
    "args": {
      "External id": 3347,"Record function id": 0, "finished": true, "Ev Idx": 3346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840911.303, "dur": 1.491,
    "args": {
      "External id": 3348,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1826, "finished": true, "Ev Idx": 3347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840912.026, "dur": 0.294,
    "args": {
      "External id": 3349,"Record function id": 0, "finished": true, "Ev Idx": 3348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840913.512, "dur": 1.765,
    "args": {
      "External id": 3350,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1827, "finished": true, "Ev Idx": 3349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840914.380, "dur": 0.440,
    "args": {
      "External id": 3351,"Record function id": 0, "finished": true, "Ev Idx": 3350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564840916.976, "dur": 2.216,
    "args": {
      "External id": 3352,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1828, "finished": true, "Ev Idx": 3351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840918.420, "dur": 0.331,
    "args": {
      "External id": 3353,"Record function id": 0, "finished": true, "Ev Idx": 3352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564840920.634, "dur": 11.825,
    "args": {
      "External id": 3354,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1829, "finished": true, "Ev Idx": 3353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564840921.502, "dur": 9.459,
    "args": {
      "External id": 3355,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1829, "finished": true, "Ev Idx": 3354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564840922.502, "dur": 1.938,
    "args": {
      "External id": 3356,"Record function id": 0, "finished": true, "Ev Idx": 3355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564840922.811, "dur": 1.507,
    "args": {
      "External id": 3357,"Record function id": 0, "finished": true, "Ev Idx": 3356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564840924.760, "dur": 5.877,
    "args": {
      "External id": 3358,"Record function id": 0, "finished": true, "Ev Idx": 3357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564840931.460, "dur": 0.812,
    "args": {
      "External id": 3359,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1830, "finished": true, "Ev Idx": 3358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564840934.495, "dur": 1.747,
    "args": {
      "External id": 3360,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1831, "finished": true, "Ev Idx": 3359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840935.389, "dur": 0.416,
    "args": {
      "External id": 3361,"Record function id": 0, "finished": true, "Ev Idx": 3360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564840937.775, "dur": 42.900,
    "args": {
      "External id": 3362,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1832, "finished": true, "Ev Idx": 3361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564840938.523, "dur": 1.611,
    "args": {
      "External id": 3363,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1832, "finished": true, "Ev Idx": 3362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840939.372, "dur": 0.327,
    "args": {
      "External id": 3364,"Record function id": 0, "finished": true, "Ev Idx": 3363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564840940.468, "dur": 5.143,
    "args": {
      "External id": 3365,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1833, "finished": true, "Ev Idx": 3364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564840942.192, "dur": 3.239,
    "args": {
      "External id": 3366,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1833, "finished": true, "Ev Idx": 3365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564840946.075, "dur": 2.492,
    "args": {
      "External id": 3367,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1834, "finished": true, "Ev Idx": 3366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840947.943, "dur": 0.272,
    "args": {
      "External id": 3368,"Record function id": 0, "finished": true, "Ev Idx": 3367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564840948.888, "dur": 1.844,
    "args": {
      "External id": 3369,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1835, "finished": true, "Ev Idx": 3368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564840949.763, "dur": 0.870,
    "args": {
      "External id": 3370,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1835, "finished": true, "Ev Idx": 3369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564840951.416, "dur": 27.376,
    "args": {
      "External id": 3371,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1836, "finished": true, "Ev Idx": 3370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564840962.467, "dur": 1.340,
    "args": {
      "External id": 3372,"Record function id": 0, "finished": true, "Ev Idx": 3371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564840979.343, "dur": 0.964,
    "args": {
      "External id": 3373,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1837, "finished": true, "Ev Idx": 3372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564840983.643, "dur": 10.001,
    "args": {
      "External id": 3374,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1838, "finished": true, "Ev Idx": 3373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840996.624, "dur": 1.817,
    "args": {
      "External id": 3375,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1839, "finished": true, "Ev Idx": 3374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840997.395, "dur": 0.409,
    "args": {
      "External id": 3376,"Record function id": 0, "finished": true, "Ev Idx": 3375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564840999.109, "dur": 1.433,
    "args": {
      "External id": 3377,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1839, "finished": true, "Ev Idx": 3376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564840999.654, "dur": 0.462,
    "args": {
      "External id": 3378,"Record function id": 0, "finished": true, "Ev Idx": 3377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564841001.234, "dur": 1.315,
    "args": {
      "External id": 3379,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1839, "finished": true, "Ev Idx": 3378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841001.799, "dur": 0.409,
    "args": {
      "External id": 3380,"Record function id": 0, "finished": true, "Ev Idx": 3379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564841004.120, "dur": 7.781,
    "args": {
      "External id": 3381,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1839, "finished": true, "Ev Idx": 3380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564841017.204, "dur": 13.027,
    "args": {
      "External id": 3382,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1840, "finished": true, "Ev Idx": 3381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564841018.313, "dur": 0.341,
    "args": {
      "External id": 3383,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1840, "finished": true, "Ev Idx": 3382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564841019.476, "dur": 10.463,
    "args": {
      "External id": 3384,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1840, "finished": true, "Ev Idx": 3383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564841033.103, "dur": 0.329,
    "args": {
      "External id": 3385,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1841, "finished": true, "Ev Idx": 3384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564841044.664, "dur": 0.463,
    "args": {
      "External id": 3386,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1841, "finished": true, "Ev Idx": 3385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564841046.863, "dur": 42.644,
    "args": {
      "External id": 3387,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1841, "finished": true, "Ev Idx": 3386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564841047.473, "dur": 3.260,
    "args": {
      "External id": 3388,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1841, "finished": true, "Ev Idx": 3387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841049.765, "dur": 0.451,
    "args": {
      "External id": 3389,"Record function id": 0, "finished": true, "Ev Idx": 3388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564841051.117, "dur": 3.376,
    "args": {
      "External id": 3390,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1842, "finished": true, "Ev Idx": 3389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841051.603, "dur": 2.768,
    "args": {
      "External id": 3391,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1842, "finished": true, "Ev Idx": 3390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564841054.918, "dur": 2.727,
    "args": {
      "External id": 3392,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1843, "finished": true, "Ev Idx": 3391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841056.876, "dur": 0.414,
    "args": {
      "External id": 3393,"Record function id": 0, "finished": true, "Ev Idx": 3392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564841057.952, "dur": 1.540,
    "args": {
      "External id": 3394,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1844, "finished": true, "Ev Idx": 3393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841058.219, "dur": 1.187,
    "args": {
      "External id": 3395,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1844, "finished": true, "Ev Idx": 3394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564841059.930, "dur": 27.560,
    "args": {
      "External id": 3396,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1845, "finished": true, "Ev Idx": 3395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564841069.376, "dur": 1.117,
    "args": {
      "External id": 3397,"Record function id": 0, "finished": true, "Ev Idx": 3396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564841088.057, "dur": 1.126,
    "args": {
      "External id": 3398,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1846, "finished": true, "Ev Idx": 3397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841091.671, "dur": 1.867,
    "args": {
      "External id": 3399,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1847, "finished": true, "Ev Idx": 3398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841092.589, "dur": 0.441,
    "args": {
      "External id": 3400,"Record function id": 0, "finished": true, "Ev Idx": 3399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564841096.176, "dur": 11.911,
    "args": {
      "External id": 3401,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1848, "finished": true, "Ev Idx": 3400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564841096.907, "dur": 11.047,
    "args": {
      "External id": 3402,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1848, "finished": true, "Ev Idx": 3401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564841098.029, "dur": 2.268,
    "args": {
      "External id": 3403,"Record function id": 0, "finished": true, "Ev Idx": 3402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564841098.653, "dur": 1.452,
    "args": {
      "External id": 3404,"Record function id": 0, "finished": true, "Ev Idx": 3403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564841100.659, "dur": 6.940,
    "args": {
      "External id": 3405,"Record function id": 0, "finished": true, "Ev Idx": 3404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564841113.942, "dur": 1.811,
    "args": {
      "External id": 3406,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1849, "finished": true, "Ev Idx": 3405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841114.429, "dur": 1.199,
    "args": {
      "External id": 3407,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1849, "finished": true, "Ev Idx": 3406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564841130.055, "dur": 34.254,
    "args": {
      "External id": 3408,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1850, "finished": true, "Ev Idx": 3407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564841130.746, "dur": 3.841,
    "args": {
      "External id": 3409,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1850, "finished": true, "Ev Idx": 3408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841132.522, "dur": 1.531,
    "args": {
      "External id": 3410,"Record function id": 0, "finished": true, "Ev Idx": 3409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841133.305, "dur": 0.597,
    "args": {
      "External id": 3411,"Record function id": 0, "finished": true, "Ev Idx": 3410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564841134.916, "dur": 29.229,
    "args": {
      "External id": 3412,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1851, "finished": true, "Ev Idx": 3411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564841135.383, "dur": 1.462,
    "args": {
      "External id": 3413,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1851, "finished": true, "Ev Idx": 3412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841135.777, "dur": 0.970,
    "args": {
      "External id": 3414,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1851, "finished": true, "Ev Idx": 3413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564841137.372, "dur": 24.813,
    "args": {
      "External id": 3415,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1852, "finished": true, "Ev Idx": 3414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564841147.142, "dur": 1.062,
    "args": {
      "External id": 3416,"Record function id": 0, "finished": true, "Ev Idx": 3415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564841162.768, "dur": 0.993,
    "args": {
      "External id": 3417,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1853, "finished": true, "Ev Idx": 3416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564841169.167, "dur": 9.427,
    "args": {
      "External id": 3418,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1854, "finished": true, "Ev Idx": 3417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564841187.575, "dur": 0.364,
    "args": {
      "External id": 3419,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1855, "finished": true, "Ev Idx": 3418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564841190.000, "dur": 11.245,
    "args": {
      "External id": 3420,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1855, "finished": true, "Ev Idx": 3419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564841191.644, "dur": 0.453,
    "args": {
      "External id": 3421,"Record function id": 0, "finished": true, "Ev Idx": 3420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564841192.476, "dur": 0.272,
    "args": {
      "External id": 3422,"Record function id": 0, "finished": true, "Ev Idx": 3421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564841203.807, "dur": 11.750,
    "args": {
      "External id": 3423,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1856, "finished": true, "Ev Idx": 3422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564841218.920, "dur": 10.495,
    "args": {
      "External id": 3424,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1857, "finished": true, "Ev Idx": 3423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564841231.231, "dur": 9.648,
    "args": {
      "External id": 3425,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1858, "finished": true, "Ev Idx": 3424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564841242.567, "dur": 7.770,
    "args": {
      "External id": 3426,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1859, "finished": true, "Ev Idx": 3425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564841252.953, "dur": 0.273,
    "args": {
      "External id": 3427,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1860, "finished": true, "Ev Idx": 3426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564841254.186, "dur": 8.864,
    "args": {
      "External id": 3428,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1860, "finished": true, "Ev Idx": 3427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564841277.590, "dur": 110.338,
    "args": {
      "External id": 3429,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1861, "finished": true, "Ev Idx": 3428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564841278.122, "dur": 3.025,
    "args": {
      "External id": 3430,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1861, "finished": true, "Ev Idx": 3429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841279.198, "dur": 1.338,
    "args": {
      "External id": 3431,"Record function id": 0, "finished": true, "Ev Idx": 3430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841279.845, "dur": 0.562,
    "args": {
      "External id": 3432,"Record function id": 0, "finished": true, "Ev Idx": 3431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564841281.535, "dur": 106.205,
    "args": {
      "External id": 3433,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1862, "finished": true, "Ev Idx": 3432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564841282.325, "dur": 2.058,
    "args": {
      "External id": 3434,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1862, "finished": true, "Ev Idx": 3433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841282.928, "dur": 1.331,
    "args": {
      "External id": 3435,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1862, "finished": true, "Ev Idx": 3434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564841284.697, "dur": 100.532,
    "args": {
      "External id": 3436,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1863, "finished": true, "Ev Idx": 3435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564841354.017, "dur": 1.349,
    "args": {
      "External id": 3437,"Record function id": 0, "finished": true, "Ev Idx": 3436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564841386.123, "dur": 1.281,
    "args": {
      "External id": 3438,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1864, "finished": true, "Ev Idx": 3437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564841397.358, "dur": 12.543,
    "args": {
      "External id": 3439,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1865, "finished": true, "Ev Idx": 3438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564841417.724, "dur": 79.669,
    "args": {
      "External id": 3440,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1866, "finished": true, "Ev Idx": 3439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564841418.393, "dur": 3.975,
    "args": {
      "External id": 3441,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1866, "finished": true, "Ev Idx": 3440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841419.725, "dur": 1.446,
    "args": {
      "External id": 3442,"Record function id": 0, "finished": true, "Ev Idx": 3441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841420.359, "dur": 0.678,
    "args": {
      "External id": 3443,"Record function id": 0, "finished": true, "Ev Idx": 3442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564841422.866, "dur": 74.408,
    "args": {
      "External id": 3444,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1867, "finished": true, "Ev Idx": 3443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564841423.285, "dur": 3.183,
    "args": {
      "External id": 3445,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1867, "finished": true, "Ev Idx": 3444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841423.705, "dur": 2.660,
    "args": {
      "External id": 3446,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1867, "finished": true, "Ev Idx": 3445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564841426.838, "dur": 68.379,
    "args": {
      "External id": 3447,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1868, "finished": true, "Ev Idx": 3446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564841470.510, "dur": 1.132,
    "args": {
      "External id": 3448,"Record function id": 0, "finished": true, "Ev Idx": 3447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564841495.815, "dur": 1.136,
    "args": {
      "External id": 3449,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1869, "finished": true, "Ev Idx": 3448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564841499.642, "dur": 9.338,
    "args": {
      "External id": 3450,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1870, "finished": true, "Ev Idx": 3449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564841516.189, "dur": 38.598,
    "args": {
      "External id": 3451,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1871, "finished": true, "Ev Idx": 3450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564841516.837, "dur": 2.804,
    "args": {
      "External id": 3452,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1871, "finished": true, "Ev Idx": 3451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841517.928, "dur": 1.037,
    "args": {
      "External id": 3453,"Record function id": 0, "finished": true, "Ev Idx": 3452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841518.358, "dur": 0.468,
    "args": {
      "External id": 3454,"Record function id": 0, "finished": true, "Ev Idx": 3453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564841519.967, "dur": 34.732,
    "args": {
      "External id": 3455,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1872, "finished": true, "Ev Idx": 3454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564841520.423, "dur": 2.045,
    "args": {
      "External id": 3456,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1872, "finished": true, "Ev Idx": 3455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841520.711, "dur": 1.641,
    "args": {
      "External id": 3457,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1872, "finished": true, "Ev Idx": 3456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564841522.792, "dur": 30.065,
    "args": {
      "External id": 3458,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1873, "finished": true, "Ev Idx": 3457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564841531.862, "dur": 1.065,
    "args": {
      "External id": 3459,"Record function id": 0, "finished": true, "Ev Idx": 3458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564841553.465, "dur": 0.957,
    "args": {
      "External id": 3460,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1874, "finished": true, "Ev Idx": 3459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564841557.635, "dur": 9.508,
    "args": {
      "External id": 3461,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1875, "finished": true, "Ev Idx": 3460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564841587.323, "dur": 0.305,
    "args": {
      "External id": 3462,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1876, "finished": true, "Ev Idx": 3461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564841589.479, "dur": 10.116,
    "args": {
      "External id": 3463,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1876, "finished": true, "Ev Idx": 3462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564841590.735, "dur": 0.458,
    "args": {
      "External id": 3464,"Record function id": 0, "finished": true, "Ev Idx": 3463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564841591.545, "dur": 0.088,
    "args": {
      "External id": 3465,"Record function id": 0, "finished": true, "Ev Idx": 3464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564841602.202, "dur": 11.254,
    "args": {
      "External id": 3466,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1877, "finished": true, "Ev Idx": 3465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564841617.009, "dur": 8.375,
    "args": {
      "External id": 3467,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1878, "finished": true, "Ev Idx": 3466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564841627.448, "dur": 9.285,
    "args": {
      "External id": 3468,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1879, "finished": true, "Ev Idx": 3467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564841638.449, "dur": 7.171,
    "args": {
      "External id": 3469,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1880, "finished": true, "Ev Idx": 3468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564841648.037, "dur": 0.079,
    "args": {
      "External id": 3470,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1881, "finished": true, "Ev Idx": 3469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564841648.948, "dur": 10.087,
    "args": {
      "External id": 3471,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1881, "finished": true, "Ev Idx": 3470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564841677.795, "dur": 36.613,
    "args": {
      "External id": 3472,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1882, "finished": true, "Ev Idx": 3471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564841678.392, "dur": 3.894,
    "args": {
      "External id": 3473,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1882, "finished": true, "Ev Idx": 3472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841679.782, "dur": 1.559,
    "args": {
      "External id": 3474,"Record function id": 0, "finished": true, "Ev Idx": 3473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841680.667, "dur": 0.533,
    "args": {
      "External id": 3475,"Record function id": 0, "finished": true, "Ev Idx": 3474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564841682.635, "dur": 31.680,
    "args": {
      "External id": 3476,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1883, "finished": true, "Ev Idx": 3475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564841683.413, "dur": 2.354,
    "args": {
      "External id": 3477,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1883, "finished": true, "Ev Idx": 3476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841683.728, "dur": 1.913,
    "args": {
      "External id": 3478,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1883, "finished": true, "Ev Idx": 3477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564841686.098, "dur": 26.141,
    "args": {
      "External id": 3479,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1884, "finished": true, "Ev Idx": 3478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564841695.715, "dur": 1.062,
    "args": {
      "External id": 3480,"Record function id": 0, "finished": true, "Ev Idx": 3479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564841712.821, "dur": 1.167,
    "args": {
      "External id": 3481,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1885, "finished": true, "Ev Idx": 3480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841717.542, "dur": 1.319,
    "args": {
      "External id": 3482,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1886, "finished": true, "Ev Idx": 3481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841720.689, "dur": 2.664,
    "args": {
      "External id": 3483,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1887, "finished": true, "Ev Idx": 3482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841721.732, "dur": 0.810,
    "args": {
      "External id": 3484,"Record function id": 0, "finished": true, "Ev Idx": 3483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564841729.914, "dur": 39.611,
    "args": {
      "External id": 3485,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1888, "finished": true, "Ev Idx": 3484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564841730.250, "dur": 2.461,
    "args": {
      "External id": 3486,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1888, "finished": true, "Ev Idx": 3485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841731.240, "dur": 1.045,
    "args": {
      "External id": 3487,"Record function id": 0, "finished": true, "Ev Idx": 3486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841731.740, "dur": 0.422,
    "args": {
      "External id": 3488,"Record function id": 0, "finished": true, "Ev Idx": 3487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564841733.041, "dur": 36.386,
    "args": {
      "External id": 3489,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1889, "finished": true, "Ev Idx": 3488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564841733.639, "dur": 1.514,
    "args": {
      "External id": 3490,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1889, "finished": true, "Ev Idx": 3489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841734.054, "dur": 1.002,
    "args": {
      "External id": 3491,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1889, "finished": true, "Ev Idx": 3490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564841735.452, "dur": 32.003,
    "args": {
      "External id": 3492,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1890, "finished": true, "Ev Idx": 3491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564841745.106, "dur": 1.056,
    "args": {
      "External id": 3493,"Record function id": 0, "finished": true, "Ev Idx": 3492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564841768.202, "dur": 0.940,
    "args": {
      "External id": 3494,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1891, "finished": true, "Ev Idx": 3493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841771.937, "dur": 1.158,
    "args": {
      "External id": 3495,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1892, "finished": true, "Ev Idx": 3494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841774.549, "dur": 1.991,
    "args": {
      "External id": 3496,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1893, "finished": true, "Ev Idx": 3495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841775.475, "dur": 0.467,
    "args": {
      "External id": 3497,"Record function id": 0, "finished": true, "Ev Idx": 3496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564841782.786, "dur": 35.280,
    "args": {
      "External id": 3498,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1894, "finished": true, "Ev Idx": 3497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564841783.091, "dur": 2.319,
    "args": {
      "External id": 3499,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1894, "finished": true, "Ev Idx": 3498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841784.023, "dur": 0.959,
    "args": {
      "External id": 3500,"Record function id": 0, "finished": true, "Ev Idx": 3499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841784.622, "dur": 0.261,
    "args": {
      "External id": 3501,"Record function id": 0, "finished": true, "Ev Idx": 3500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564841785.686, "dur": 32.274,
    "args": {
      "External id": 3502,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1895, "finished": true, "Ev Idx": 3501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564841786.070, "dur": 1.411,
    "args": {
      "External id": 3503,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1895, "finished": true, "Ev Idx": 3502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841786.535, "dur": 0.858,
    "args": {
      "External id": 3504,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1895, "finished": true, "Ev Idx": 3503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564841787.806, "dur": 28.385,
    "args": {
      "External id": 3505,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1896, "finished": true, "Ev Idx": 3504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564841796.423, "dur": 1.213,
    "args": {
      "External id": 3506,"Record function id": 0, "finished": true, "Ev Idx": 3505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564841816.716, "dur": 0.978,
    "args": {
      "External id": 3507,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1897, "finished": true, "Ev Idx": 3506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564841820.398, "dur": 1.332,
    "args": {
      "External id": 3508,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1898, "finished": true, "Ev Idx": 3507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564841823.149, "dur": 1.659,
    "args": {
      "External id": 3509,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1899, "finished": true, "Ev Idx": 3508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841823.929, "dur": 0.394,
    "args": {
      "External id": 3510,"Record function id": 0, "finished": true, "Ev Idx": 3509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564841827.557, "dur": 2.156,
    "args": {
      "External id": 3511,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1900, "finished": true, "Ev Idx": 3510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841828.784, "dur": 0.333,
    "args": {
      "External id": 3512,"Record function id": 0, "finished": true, "Ev Idx": 3511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564841830.766, "dur": 1.135,
    "args": {
      "External id": 3513,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1900, "finished": true, "Ev Idx": 3512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841831.237, "dur": 0.319,
    "args": {
      "External id": 3514,"Record function id": 0, "finished": true, "Ev Idx": 3513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564841832.817, "dur": 46.633,
    "args": {
      "External id": 3515,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1900, "finished": true, "Ev Idx": 3514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564841885.906, "dur": 5.094,
    "args": {
      "External id": 3516,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1901, "finished": true, "Ev Idx": 3515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841888.779, "dur": 0.697,
    "args": {
      "External id": 3517,"Record function id": 0, "finished": true, "Ev Idx": 3516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564841893.330, "dur": 1.866,
    "args": {
      "External id": 3518,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1902, "finished": true, "Ev Idx": 3517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841894.228, "dur": 0.362,
    "args": {
      "External id": 3519,"Record function id": 0, "finished": true, "Ev Idx": 3518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564841897.090, "dur": 12.830,
    "args": {
      "External id": 3520,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1903, "finished": true, "Ev Idx": 3519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564841914.399, "dur": 17.637,
    "args": {
      "External id": 3521,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1904, "finished": true, "Ev Idx": 3520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564841935.209, "dur": 9.108,
    "args": {
      "External id": 3522,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1905, "finished": true, "Ev Idx": 3521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564841945.958, "dur": 7.742,
    "args": {
      "External id": 3523,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1906, "finished": true, "Ev Idx": 3522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564841955.789, "dur": 7.571,
    "args": {
      "External id": 3524,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1907, "finished": true, "Ev Idx": 3523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564841966.124, "dur": 2.890,
    "args": {
      "External id": 3525,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1908, "finished": true, "Ev Idx": 3524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841967.830, "dur": 0.514,
    "args": {
      "External id": 3526,"Record function id": 0, "finished": true, "Ev Idx": 3525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564841970.561, "dur": 2.309,
    "args": {
      "External id": 3527,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1909, "finished": true, "Ev Idx": 3526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564841971.986, "dur": 0.330,
    "args": {
      "External id": 3528,"Record function id": 0, "finished": true, "Ev Idx": 3527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564841973.932, "dur": 7.543,
    "args": {
      "External id": 3529,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1910, "finished": true, "Ev Idx": 3528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564841983.442, "dur": 12.210,
    "args": {
      "External id": 3530,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1911, "finished": true, "Ev Idx": 3529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564841997.808, "dur": 6.667,
    "args": {
      "External id": 3531,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1912, "finished": true, "Ev Idx": 3530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564842005.906, "dur": 6.779,
    "args": {
      "External id": 3532,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1913, "finished": true, "Ev Idx": 3531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842037.082, "dur": 3.434,
    "args": {
      "External id": 3533,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1914, "finished": true, "Ev Idx": 3532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842039.257, "dur": 0.483,
    "args": {
      "External id": 3534,"Record function id": 0, "finished": true, "Ev Idx": 3533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842041.168, "dur": 2.262,
    "args": {
      "External id": 3535,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1915, "finished": true, "Ev Idx": 3534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842042.670, "dur": 0.314,
    "args": {
      "External id": 3536,"Record function id": 0, "finished": true, "Ev Idx": 3535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564842044.172, "dur": 2.482,
    "args": {
      "External id": 3537,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1916, "finished": true, "Ev Idx": 3536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842045.090, "dur": 0.892,
    "args": {
      "External id": 3538,"Record function id": 0, "finished": true, "Ev Idx": 3537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842047.483, "dur": 1.844,
    "args": {
      "External id": 3539,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1917, "finished": true, "Ev Idx": 3538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842048.433, "dur": 0.372,
    "args": {
      "External id": 3540,"Record function id": 0, "finished": true, "Ev Idx": 3539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842050.125, "dur": 1.999,
    "args": {
      "External id": 3541,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1918, "finished": true, "Ev Idx": 3540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842051.184, "dur": 0.487,
    "args": {
      "External id": 3542,"Record function id": 0, "finished": true, "Ev Idx": 3541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564842055.138, "dur": 2.944,
    "args": {
      "External id": 3543,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1919, "finished": true, "Ev Idx": 3542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842056.949, "dur": 0.463,
    "args": {
      "External id": 3544,"Record function id": 0, "finished": true, "Ev Idx": 3543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842060.540, "dur": 45.258,
    "args": {
      "External id": 3545,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1920, "finished": true, "Ev Idx": 3544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564842063.098, "dur": 38.834,
    "args": {
      "External id": 3546,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1920, "finished": true, "Ev Idx": 3545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564842065.482, "dur": 21.039,
    "args": {
      "External id": 3547,"Record function id": 0, "finished": true, "Ev Idx": 3546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564842066.856, "dur": 1.861,
    "args": {
      "External id": 3548,"Record function id": 0, "finished": true, "Ev Idx": 3547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564842088.163, "dur": 12.764,
    "args": {
      "External id": 3549,"Record function id": 0, "finished": true, "Ev Idx": 3548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564842103.007, "dur": 2.383,
    "args": {
      "External id": 3550,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1921, "finished": true, "Ev Idx": 3549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842110.519, "dur": 3.018,
    "args": {
      "External id": 3551,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1922, "finished": true, "Ev Idx": 3550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842112.304, "dur": 0.556,
    "args": {
      "External id": 3552,"Record function id": 0, "finished": true, "Ev Idx": 3551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842114.249, "dur": 1.551,
    "args": {
      "External id": 3553,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1923, "finished": true, "Ev Idx": 3552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842114.988, "dur": 0.314,
    "args": {
      "External id": 3554,"Record function id": 0, "finished": true, "Ev Idx": 3553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564842116.381, "dur": 1.899,
    "args": {
      "External id": 3555,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1924, "finished": true, "Ev Idx": 3554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842117.339, "dur": 0.506,
    "args": {
      "External id": 3556,"Record function id": 0, "finished": true, "Ev Idx": 3555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842118.861, "dur": 1.772,
    "args": {
      "External id": 3557,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1925, "finished": true, "Ev Idx": 3556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842119.826, "dur": 0.323,
    "args": {
      "External id": 3558,"Record function id": 0, "finished": true, "Ev Idx": 3557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842121.252, "dur": 2.677,
    "args": {
      "External id": 3559,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1926, "finished": true, "Ev Idx": 3558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842123.041, "dur": 0.459,
    "args": {
      "External id": 3560,"Record function id": 0, "finished": true, "Ev Idx": 3559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564842125.845, "dur": 2.190,
    "args": {
      "External id": 3561,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1927, "finished": true, "Ev Idx": 3560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842127.190, "dur": 0.372,
    "args": {
      "External id": 3562,"Record function id": 0, "finished": true, "Ev Idx": 3561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842129.667, "dur": 13.446,
    "args": {
      "External id": 3563,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1928, "finished": true, "Ev Idx": 3562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564842131.216, "dur": 10.329,
    "args": {
      "External id": 3564,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1928, "finished": true, "Ev Idx": 3563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564842132.048, "dur": 2.906,
    "args": {
      "External id": 3565,"Record function id": 0, "finished": true, "Ev Idx": 3564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564842132.670, "dur": 2.118,
    "args": {
      "External id": 3566,"Record function id": 0, "finished": true, "Ev Idx": 3565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564842135.268, "dur": 5.911,
    "args": {
      "External id": 3567,"Record function id": 0, "finished": true, "Ev Idx": 3566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564842141.954, "dur": 0.971,
    "args": {
      "External id": 3568,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1929, "finished": true, "Ev Idx": 3567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564842145.404, "dur": 1.941,
    "args": {
      "External id": 3569,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1930, "finished": true, "Ev Idx": 3568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842146.424, "dur": 0.377,
    "args": {
      "External id": 3570,"Record function id": 0, "finished": true, "Ev Idx": 3569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564842149.101, "dur": 55.141,
    "args": {
      "External id": 3571,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1931, "finished": true, "Ev Idx": 3570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564842149.999, "dur": 3.124,
    "args": {
      "External id": 3572,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1931, "finished": true, "Ev Idx": 3571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842152.170, "dur": 0.545,
    "args": {
      "External id": 3573,"Record function id": 0, "finished": true, "Ev Idx": 3572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842153.520, "dur": 4.891,
    "args": {
      "External id": 3574,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1932, "finished": true, "Ev Idx": 3573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564842155.514, "dur": 2.738,
    "args": {
      "External id": 3575,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1932, "finished": true, "Ev Idx": 3574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564842159.084, "dur": 1.524,
    "args": {
      "External id": 3576,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1933, "finished": true, "Ev Idx": 3575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842159.984, "dur": 0.288,
    "args": {
      "External id": 3577,"Record function id": 0, "finished": true, "Ev Idx": 3576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842161.039, "dur": 1.707,
    "args": {
      "External id": 3578,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1934, "finished": true, "Ev Idx": 3577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564842161.806, "dur": 0.844,
    "args": {
      "External id": 3579,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1934, "finished": true, "Ev Idx": 3578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564842163.447, "dur": 38.742,
    "args": {
      "External id": 3580,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1935, "finished": true, "Ev Idx": 3579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564842178.559, "dur": 1.177,
    "args": {
      "External id": 3581,"Record function id": 0, "finished": true, "Ev Idx": 3580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564842202.814, "dur": 1.022,
    "args": {
      "External id": 3582,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1936, "finished": true, "Ev Idx": 3581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564842208.334, "dur": 10.161,
    "args": {
      "External id": 3583,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1937, "finished": true, "Ev Idx": 3582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842221.585, "dur": 39.207,
    "args": {
      "External id": 3584,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1938, "finished": true, "Ev Idx": 3583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842259.738, "dur": 0.411,
    "args": {
      "External id": 3585,"Record function id": 0, "finished": true, "Ev Idx": 3584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842261.537, "dur": 1.300,
    "args": {
      "External id": 3586,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1938, "finished": true, "Ev Idx": 3585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842262.126, "dur": 0.367,
    "args": {
      "External id": 3587,"Record function id": 0, "finished": true, "Ev Idx": 3586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564842263.490, "dur": 1.082,
    "args": {
      "External id": 3588,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1938, "finished": true, "Ev Idx": 3587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842263.965, "dur": 0.295,
    "args": {
      "External id": 3589,"Record function id": 0, "finished": true, "Ev Idx": 3588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564842266.322, "dur": 9.782,
    "args": {
      "External id": 3590,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1938, "finished": true, "Ev Idx": 3589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564842281.466, "dur": 14.557,
    "args": {
      "External id": 3591,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1939, "finished": true, "Ev Idx": 3590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564842282.474, "dur": 0.383,
    "args": {
      "External id": 3592,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1939, "finished": true, "Ev Idx": 3591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564842283.515, "dur": 12.243,
    "args": {
      "External id": 3593,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1939, "finished": true, "Ev Idx": 3592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564842299.097, "dur": 0.191,
    "args": {
      "External id": 3594,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1940, "finished": true, "Ev Idx": 3593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564842310.197, "dur": 0.476,
    "args": {
      "External id": 3595,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1940, "finished": true, "Ev Idx": 3594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564842312.353, "dur": 40.621,
    "args": {
      "External id": 3596,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1940, "finished": true, "Ev Idx": 3595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564842312.883, "dur": 2.704,
    "args": {
      "External id": 3597,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1940, "finished": true, "Ev Idx": 3596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842314.717, "dur": 0.380,
    "args": {
      "External id": 3598,"Record function id": 0, "finished": true, "Ev Idx": 3597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842315.921, "dur": 2.262,
    "args": {
      "External id": 3599,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1941, "finished": true, "Ev Idx": 3598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564842316.248, "dur": 1.793,
    "args": {
      "External id": 3600,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1941, "finished": true, "Ev Idx": 3599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564842318.616, "dur": 2.725,
    "args": {
      "External id": 3601,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1942, "finished": true, "Ev Idx": 3600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842320.735, "dur": 0.243,
    "args": {
      "External id": 3602,"Record function id": 0, "finished": true, "Ev Idx": 3601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842321.626, "dur": 1.824,
    "args": {
      "External id": 3603,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1943, "finished": true, "Ev Idx": 3602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564842321.897, "dur": 1.478,
    "args": {
      "External id": 3604,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1943, "finished": true, "Ev Idx": 3603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564842323.809, "dur": 27.150,
    "args": {
      "External id": 3605,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1944, "finished": true, "Ev Idx": 3604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564842333.065, "dur": 0.993,
    "args": {
      "External id": 3606,"Record function id": 0, "finished": true, "Ev Idx": 3605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564842351.533, "dur": 1.140,
    "args": {
      "External id": 3607,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1945, "finished": true, "Ev Idx": 3606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564842355.012, "dur": 1.733,
    "args": {
      "External id": 3608,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1946, "finished": true, "Ev Idx": 3607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842355.841, "dur": 0.372,
    "args": {
      "External id": 3609,"Record function id": 0, "finished": true, "Ev Idx": 3608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564842359.420, "dur": 10.666,
    "args": {
      "External id": 3610,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1947, "finished": true, "Ev Idx": 3609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564842360.001, "dur": 9.947,
    "args": {
      "External id": 3611,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1947, "finished": true, "Ev Idx": 3610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564842361.132, "dur": 1.756,
    "args": {
      "External id": 3612,"Record function id": 0, "finished": true, "Ev Idx": 3611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564842361.551, "dur": 1.164,
    "args": {
      "External id": 3613,"Record function id": 0, "finished": true, "Ev Idx": 3612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564842363.277, "dur": 6.286,
    "args": {
      "External id": 3614,"Record function id": 0, "finished": true, "Ev Idx": 3613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842376.007, "dur": 1.706,
    "args": {
      "External id": 3615,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1948, "finished": true, "Ev Idx": 3614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564842376.337, "dur": 1.263,
    "args": {
      "External id": 3616,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1948, "finished": true, "Ev Idx": 3615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564842393.477, "dur": 36.512,
    "args": {
      "External id": 3617,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1949, "finished": true, "Ev Idx": 3616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564842394.070, "dur": 3.007,
    "args": {
      "External id": 3618,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1949, "finished": true, "Ev Idx": 3617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564842395.633, "dur": 0.859,
    "args": {
      "External id": 3619,"Record function id": 0, "finished": true, "Ev Idx": 3618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842395.994, "dur": 0.379,
    "args": {
      "External id": 3620,"Record function id": 0, "finished": true, "Ev Idx": 3619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564842397.456, "dur": 32.332,
    "args": {
      "External id": 3621,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1950, "finished": true, "Ev Idx": 3620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842397.912, "dur": 1.926,
    "args": {
      "External id": 3622,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1950, "finished": true, "Ev Idx": 3621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564842398.185, "dur": 1.549,
    "args": {
      "External id": 3623,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1950, "finished": true, "Ev Idx": 3622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564842400.376, "dur": 27.567,
    "args": {
      "External id": 3624,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1951, "finished": true, "Ev Idx": 3623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564842411.182, "dur": 1.052,
    "args": {
      "External id": 3625,"Record function id": 0, "finished": true, "Ev Idx": 3624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564842428.516, "dur": 0.923,
    "args": {
      "External id": 3626,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1952, "finished": true, "Ev Idx": 3625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564842435.169, "dur": 8.562,
    "args": {
      "External id": 3627,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1953, "finished": true, "Ev Idx": 3626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564842452.884, "dur": 0.145,
    "args": {
      "External id": 3628,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1954, "finished": true, "Ev Idx": 3627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564842455.120, "dur": 10.462,
    "args": {
      "External id": 3629,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1954, "finished": true, "Ev Idx": 3628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564842456.942, "dur": 0.390,
    "args": {
      "External id": 3630,"Record function id": 0, "finished": true, "Ev Idx": 3629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564842457.745, "dur": 0.076,
    "args": {
      "External id": 3631,"Record function id": 0, "finished": true, "Ev Idx": 3630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564842468.593, "dur": 12.718,
    "args": {
      "External id": 3632,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1955, "finished": true, "Ev Idx": 3631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564842484.897, "dur": 8.351,
    "args": {
      "External id": 3633,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1956, "finished": true, "Ev Idx": 3632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564842495.345, "dur": 10.466,
    "args": {
      "External id": 3634,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1957, "finished": true, "Ev Idx": 3633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564842507.666, "dur": 9.143,
    "args": {
      "External id": 3635,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1958, "finished": true, "Ev Idx": 3634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564842519.309, "dur": 0.140,
    "args": {
      "External id": 3636,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1959, "finished": true, "Ev Idx": 3635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564842520.414, "dur": 8.346,
    "args": {
      "External id": 3637,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1959, "finished": true, "Ev Idx": 3636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564842542.083, "dur": 48.361,
    "args": {
      "External id": 3638,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1960, "finished": true, "Ev Idx": 3637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564842542.503, "dur": 2.986,
    "args": {
      "External id": 3639,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1960, "finished": true, "Ev Idx": 3638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564842543.623, "dur": 1.245,
    "args": {
      "External id": 3640,"Record function id": 0, "finished": true, "Ev Idx": 3639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842544.293, "dur": 0.436,
    "args": {
      "External id": 3641,"Record function id": 0, "finished": true, "Ev Idx": 3640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564842545.852, "dur": 44.483,
    "args": {
      "External id": 3642,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1961, "finished": true, "Ev Idx": 3641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842546.325, "dur": 1.837,
    "args": {
      "External id": 3643,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1961, "finished": true, "Ev Idx": 3642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564842546.643, "dur": 1.386,
    "args": {
      "External id": 3644,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1961, "finished": true, "Ev Idx": 3643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564842548.487, "dur": 40.000,
    "args": {
      "External id": 3645,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1962, "finished": true, "Ev Idx": 3644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564842562.020, "dur": 1.202,
    "args": {
      "External id": 3646,"Record function id": 0, "finished": true, "Ev Idx": 3645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564842589.095, "dur": 0.914,
    "args": {
      "External id": 3647,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1963, "finished": true, "Ev Idx": 3646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564842599.047, "dur": 70.551,
    "args": {
      "External id": 3648,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1964, "finished": true, "Ev Idx": 3647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564842678.200, "dur": 45.557,
    "args": {
      "External id": 3649,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1965, "finished": true, "Ev Idx": 3648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564842678.728, "dur": 4.685,
    "args": {
      "External id": 3650,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1965, "finished": true, "Ev Idx": 3649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564842681.015, "dur": 1.569,
    "args": {
      "External id": 3651,"Record function id": 0, "finished": true, "Ev Idx": 3650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842681.919, "dur": 0.506,
    "args": {
      "External id": 3652,"Record function id": 0, "finished": true, "Ev Idx": 3651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564842683.791, "dur": 39.820,
    "args": {
      "External id": 3653,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1966, "finished": true, "Ev Idx": 3652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842684.286, "dur": 1.688,
    "args": {
      "External id": 3654,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1966, "finished": true, "Ev Idx": 3653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564842684.593, "dur": 1.270,
    "args": {
      "External id": 3655,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1966, "finished": true, "Ev Idx": 3654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564842686.344, "dur": 35.363,
    "args": {
      "External id": 3656,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1967, "finished": true, "Ev Idx": 3655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564842697.158, "dur": 1.174,
    "args": {
      "External id": 3657,"Record function id": 0, "finished": true, "Ev Idx": 3656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564842722.278, "dur": 0.994,
    "args": {
      "External id": 3658,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1968, "finished": true, "Ev Idx": 3657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564842725.587, "dur": 42.499,
    "args": {
      "External id": 3659,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1969, "finished": true, "Ev Idx": 3658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564842775.260, "dur": 40.281,
    "args": {
      "External id": 3660,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1970, "finished": true, "Ev Idx": 3659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564842775.672, "dur": 3.131,
    "args": {
      "External id": 3661,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1970, "finished": true, "Ev Idx": 3660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564842776.807, "dur": 1.181,
    "args": {
      "External id": 3662,"Record function id": 0, "finished": true, "Ev Idx": 3661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842777.398, "dur": 0.463,
    "args": {
      "External id": 3663,"Record function id": 0, "finished": true, "Ev Idx": 3662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564842779.134, "dur": 36.303,
    "args": {
      "External id": 3664,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1971, "finished": true, "Ev Idx": 3663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842779.594, "dur": 2.641,
    "args": {
      "External id": 3665,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1971, "finished": true, "Ev Idx": 3664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564842779.879, "dur": 2.261,
    "args": {
      "External id": 3666,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1971, "finished": true, "Ev Idx": 3665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564842782.579, "dur": 31.024,
    "args": {
      "External id": 3667,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1972, "finished": true, "Ev Idx": 3666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564842791.936, "dur": 1.109,
    "args": {
      "External id": 3668,"Record function id": 0, "finished": true, "Ev Idx": 3667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564842814.198, "dur": 0.939,
    "args": {
      "External id": 3669,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1973, "finished": true, "Ev Idx": 3668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564842818.692, "dur": 8.953,
    "args": {
      "External id": 3670,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1974, "finished": true, "Ev Idx": 3669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564842848.826, "dur": 0.214,
    "args": {
      "External id": 3671,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1975, "finished": true, "Ev Idx": 3670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564842851.047, "dur": 11.231,
    "args": {
      "External id": 3672,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1975, "finished": true, "Ev Idx": 3671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564842852.404, "dur": 0.489,
    "args": {
      "External id": 3673,"Record function id": 0, "finished": true, "Ev Idx": 3672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564842853.219, "dur": 0.068,
    "args": {
      "External id": 3674,"Record function id": 0, "finished": true, "Ev Idx": 3673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564842864.911, "dur": 12.741,
    "args": {
      "External id": 3675,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1976, "finished": true, "Ev Idx": 3674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564842881.464, "dur": 8.242,
    "args": {
      "External id": 3676,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1977, "finished": true, "Ev Idx": 3675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564842891.518, "dur": 9.047,
    "args": {
      "External id": 3677,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1978, "finished": true, "Ev Idx": 3676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564842902.097, "dur": 6.982,
    "args": {
      "External id": 3678,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1979, "finished": true, "Ev Idx": 3677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564842911.587, "dur": 0.095,
    "args": {
      "External id": 3679,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1980, "finished": true, "Ev Idx": 3678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564842912.532, "dur": 8.366,
    "args": {
      "External id": 3680,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1980, "finished": true, "Ev Idx": 3679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564842939.242, "dur": 35.640,
    "args": {
      "External id": 3681,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1981, "finished": true, "Ev Idx": 3680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564842939.692, "dur": 3.631,
    "args": {
      "External id": 3682,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1981, "finished": true, "Ev Idx": 3681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564842940.739, "dur": 1.491,
    "args": {
      "External id": 3683,"Record function id": 0, "finished": true, "Ev Idx": 3682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842941.506, "dur": 0.586,
    "args": {
      "External id": 3684,"Record function id": 0, "finished": true, "Ev Idx": 3683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564842943.743, "dur": 31.040,
    "args": {
      "External id": 3685,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1982, "finished": true, "Ev Idx": 3684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842944.245, "dur": 2.356,
    "args": {
      "External id": 3686,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1982, "finished": true, "Ev Idx": 3685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564842944.578, "dur": 1.890,
    "args": {
      "External id": 3687,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1982, "finished": true, "Ev Idx": 3686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564842946.969, "dur": 25.825,
    "args": {
      "External id": 3688,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1983, "finished": true, "Ev Idx": 3687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564842956.674, "dur": 1.091,
    "args": {
      "External id": 3689,"Record function id": 0, "finished": true, "Ev Idx": 3688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564842973.399, "dur": 1.043,
    "args": {
      "External id": 3690,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1984, "finished": true, "Ev Idx": 3689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564842978.154, "dur": 1.422,
    "args": {
      "External id": 3691,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1985, "finished": true, "Ev Idx": 3690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564842981.254, "dur": 2.253,
    "args": {
      "External id": 3692,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1986, "finished": true, "Ev Idx": 3691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842982.123, "dur": 0.605,
    "args": {
      "External id": 3693,"Record function id": 0, "finished": true, "Ev Idx": 3692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564842990.570, "dur": 39.866,
    "args": {
      "External id": 3694,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1987, "finished": true, "Ev Idx": 3693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564842990.918, "dur": 2.243,
    "args": {
      "External id": 3695,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1987, "finished": true, "Ev Idx": 3694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564842991.794, "dur": 0.909,
    "args": {
      "External id": 3696,"Record function id": 0, "finished": true, "Ev Idx": 3695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564842992.307, "dur": 0.285,
    "args": {
      "External id": 3697,"Record function id": 0, "finished": true, "Ev Idx": 3696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564842993.454, "dur": 36.883,
    "args": {
      "External id": 3698,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1988, "finished": true, "Ev Idx": 3697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564842993.867, "dur": 1.447,
    "args": {
      "External id": 3699,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1988, "finished": true, "Ev Idx": 3698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564842994.139, "dur": 1.072,
    "args": {
      "External id": 3700,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1988, "finished": true, "Ev Idx": 3699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564842995.608, "dur": 32.933,
    "args": {
      "External id": 3701,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1989, "finished": true, "Ev Idx": 3700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843004.716, "dur": 1.116,
    "args": {
      "External id": 3702,"Record function id": 0, "finished": true, "Ev Idx": 3701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564843029.056, "dur": 0.976,
    "args": {
      "External id": 3703,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1990, "finished": true, "Ev Idx": 3702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564843032.792, "dur": 1.285,
    "args": {
      "External id": 3704,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1991, "finished": true, "Ev Idx": 3703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564843035.452, "dur": 1.942,
    "args": {
      "External id": 3705,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1992, "finished": true, "Ev Idx": 3704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843036.338, "dur": 0.468,
    "args": {
      "External id": 3706,"Record function id": 0, "finished": true, "Ev Idx": 3705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564843044.081, "dur": 37.294,
    "args": {
      "External id": 3707,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1993, "finished": true, "Ev Idx": 3706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564843044.360, "dur": 1.993,
    "args": {
      "External id": 3708,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1993, "finished": true, "Ev Idx": 3707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564843045.109, "dur": 0.815,
    "args": {
      "External id": 3709,"Record function id": 0, "finished": true, "Ev Idx": 3708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843045.543, "dur": 0.266,
    "args": {
      "External id": 3710,"Record function id": 0, "finished": true, "Ev Idx": 3709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564843046.682, "dur": 34.596,
    "args": {
      "External id": 3711,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1994, "finished": true, "Ev Idx": 3710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843047.097, "dur": 1.195,
    "args": {
      "External id": 3712,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1994, "finished": true, "Ev Idx": 3711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564843047.370, "dur": 0.827,
    "args": {
      "External id": 3713,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1994, "finished": true, "Ev Idx": 3712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564843048.588, "dur": 30.932,
    "args": {
      "External id": 3714,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1995, "finished": true, "Ev Idx": 3713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843058.597, "dur": 1.223,
    "args": {
      "External id": 3715,"Record function id": 0, "finished": true, "Ev Idx": 3714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564843080.036, "dur": 0.941,
    "args": {
      "External id": 3716,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1996, "finished": true, "Ev Idx": 3715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564843083.793, "dur": 1.313,
    "args": {
      "External id": 3717,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1997, "finished": true, "Ev Idx": 3716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564843086.577, "dur": 1.724,
    "args": {
      "External id": 3718,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1998, "finished": true, "Ev Idx": 3717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843087.472, "dur": 0.337,
    "args": {
      "External id": 3719,"Record function id": 0, "finished": true, "Ev Idx": 3718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564843091.056, "dur": 2.187,
    "args": {
      "External id": 3720,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1999, "finished": true, "Ev Idx": 3719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843092.462, "dur": 0.288,
    "args": {
      "External id": 3721,"Record function id": 0, "finished": true, "Ev Idx": 3720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564843094.854, "dur": 1.406,
    "args": {
      "External id": 3722,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1999, "finished": true, "Ev Idx": 3721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843095.574, "dur": 0.338,
    "args": {
      "External id": 3723,"Record function id": 0, "finished": true, "Ev Idx": 3722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564843097.252, "dur": 10.035,
    "args": {
      "External id": 3724,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 1999, "finished": true, "Ev Idx": 3723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843112.539, "dur": 5.009,
    "args": {
      "External id": 3725,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2000, "finished": true, "Ev Idx": 3724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843114.887, "dur": 1.012,
    "args": {
      "External id": 3726,"Record function id": 0, "finished": true, "Ev Idx": 3725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843119.698, "dur": 3.547,
    "args": {
      "External id": 3727,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2001, "finished": true, "Ev Idx": 3726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843121.460, "dur": 0.847,
    "args": {
      "External id": 3728,"Record function id": 0, "finished": true, "Ev Idx": 3727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564843125.124, "dur": 10.233,
    "args": {
      "External id": 3729,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2002, "finished": true, "Ev Idx": 3728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564843139.598, "dur": 18.052,
    "args": {
      "External id": 3730,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2003, "finished": true, "Ev Idx": 3729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564843160.449, "dur": 7.874,
    "args": {
      "External id": 3731,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2004, "finished": true, "Ev Idx": 3730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564843169.923, "dur": 7.840,
    "args": {
      "External id": 3732,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2005, "finished": true, "Ev Idx": 3731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564843179.732, "dur": 7.334,
    "args": {
      "External id": 3733,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2006, "finished": true, "Ev Idx": 3732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843189.328, "dur": 3.076,
    "args": {
      "External id": 3734,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2007, "finished": true, "Ev Idx": 3733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843191.234, "dur": 0.464,
    "args": {
      "External id": 3735,"Record function id": 0, "finished": true, "Ev Idx": 3734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843193.949, "dur": 2.732,
    "args": {
      "External id": 3736,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2008, "finished": true, "Ev Idx": 3735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843195.793, "dur": 0.365,
    "args": {
      "External id": 3737,"Record function id": 0, "finished": true, "Ev Idx": 3736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564843197.703, "dur": 7.066,
    "args": {
      "External id": 3738,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2009, "finished": true, "Ev Idx": 3737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564843206.877, "dur": 10.993,
    "args": {
      "External id": 3739,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2010, "finished": true, "Ev Idx": 3738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564843220.005, "dur": 9.427,
    "args": {
      "External id": 3740,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2011, "finished": true, "Ev Idx": 3739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564843230.678, "dur": 7.418,
    "args": {
      "External id": 3741,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2012, "finished": true, "Ev Idx": 3740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843262.097, "dur": 2.311,
    "args": {
      "External id": 3742,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2013, "finished": true, "Ev Idx": 3741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843263.246, "dur": 0.458,
    "args": {
      "External id": 3743,"Record function id": 0, "finished": true, "Ev Idx": 3742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843265.064, "dur": 1.841,
    "args": {
      "External id": 3744,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2014, "finished": true, "Ev Idx": 3743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843266.034, "dur": 0.330,
    "args": {
      "External id": 3745,"Record function id": 0, "finished": true, "Ev Idx": 3744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564843267.894, "dur": 3.395,
    "args": {
      "External id": 3746,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2015, "finished": true, "Ev Idx": 3745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843269.856, "dur": 0.788,
    "args": {
      "External id": 3747,"Record function id": 0, "finished": true, "Ev Idx": 3746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843272.105, "dur": 3.115,
    "args": {
      "External id": 3748,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2016, "finished": true, "Ev Idx": 3747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843273.984, "dur": 0.727,
    "args": {
      "External id": 3749,"Record function id": 0, "finished": true, "Ev Idx": 3748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843275.980, "dur": 2.511,
    "args": {
      "External id": 3750,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2017, "finished": true, "Ev Idx": 3749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843277.574, "dur": 0.464,
    "args": {
      "External id": 3751,"Record function id": 0, "finished": true, "Ev Idx": 3750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564843281.304, "dur": 2.990,
    "args": {
      "External id": 3752,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2018, "finished": true, "Ev Idx": 3751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843283.175, "dur": 0.394,
    "args": {
      "External id": 3753,"Record function id": 0, "finished": true, "Ev Idx": 3752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843286.684, "dur": 18.425,
    "args": {
      "External id": 3754,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2019, "finished": true, "Ev Idx": 3753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564843289.027, "dur": 14.107,
    "args": {
      "External id": 3755,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2019, "finished": true, "Ev Idx": 3754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564843290.429, "dur": 2.573,
    "args": {
      "External id": 3756,"Record function id": 0, "finished": true, "Ev Idx": 3755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843291.208, "dur": 1.604,
    "args": {
      "External id": 3757,"Record function id": 0, "finished": true, "Ev Idx": 3756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564843293.572, "dur": 8.963,
    "args": {
      "External id": 3758,"Record function id": 0, "finished": true, "Ev Idx": 3757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564843303.630, "dur": 1.217,
    "args": {
      "External id": 3759,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2020, "finished": true, "Ev Idx": 3758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843308.283, "dur": 2.303,
    "args": {
      "External id": 3760,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2021, "finished": true, "Ev Idx": 3759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843309.528, "dur": 0.499,
    "args": {
      "External id": 3761,"Record function id": 0, "finished": true, "Ev Idx": 3760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843311.230, "dur": 2.282,
    "args": {
      "External id": 3762,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2022, "finished": true, "Ev Idx": 3761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843312.780, "dur": 0.262,
    "args": {
      "External id": 3763,"Record function id": 0, "finished": true, "Ev Idx": 3762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564843314.063, "dur": 2.480,
    "args": {
      "External id": 3764,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2023, "finished": true, "Ev Idx": 3763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843315.787, "dur": 0.332,
    "args": {
      "External id": 3765,"Record function id": 0, "finished": true, "Ev Idx": 3764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843317.085, "dur": 1.578,
    "args": {
      "External id": 3766,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2024, "finished": true, "Ev Idx": 3765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843317.905, "dur": 0.291,
    "args": {
      "External id": 3767,"Record function id": 0, "finished": true, "Ev Idx": 3766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843319.258, "dur": 1.758,
    "args": {
      "External id": 3768,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2025, "finished": true, "Ev Idx": 3767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843320.272, "dur": 0.279,
    "args": {
      "External id": 3769,"Record function id": 0, "finished": true, "Ev Idx": 3768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564843322.529, "dur": 1.726,
    "args": {
      "External id": 3770,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2026, "finished": true, "Ev Idx": 3769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843323.335, "dur": 0.498,
    "args": {
      "External id": 3771,"Record function id": 0, "finished": true, "Ev Idx": 3770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843325.708, "dur": 14.578,
    "args": {
      "External id": 3772,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2027, "finished": true, "Ev Idx": 3771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564843326.776, "dur": 11.322,
    "args": {
      "External id": 3773,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2027, "finished": true, "Ev Idx": 3772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564843328.089, "dur": 2.522,
    "args": {
      "External id": 3774,"Record function id": 0, "finished": true, "Ev Idx": 3773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843328.780, "dur": 1.691,
    "args": {
      "External id": 3775,"Record function id": 0, "finished": true, "Ev Idx": 3774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564843331.102, "dur": 6.624,
    "args": {
      "External id": 3776,"Record function id": 0, "finished": true, "Ev Idx": 3775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564843338.508, "dur": 1.588,
    "args": {
      "External id": 3777,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2028, "finished": true, "Ev Idx": 3776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564843342.429, "dur": 1.909,
    "args": {
      "External id": 3778,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2029, "finished": true, "Ev Idx": 3777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843343.296, "dur": 0.565,
    "args": {
      "External id": 3779,"Record function id": 0, "finished": true, "Ev Idx": 3778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564843345.822, "dur": 48.277,
    "args": {
      "External id": 3780,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2030, "finished": true, "Ev Idx": 3779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564843346.386, "dur": 1.585,
    "args": {
      "External id": 3781,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2030, "finished": true, "Ev Idx": 3780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843347.206, "dur": 0.312,
    "args": {
      "External id": 3782,"Record function id": 0, "finished": true, "Ev Idx": 3781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843348.365, "dur": 3.938,
    "args": {
      "External id": 3783,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2031, "finished": true, "Ev Idx": 3782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564843350.184, "dur": 1.932,
    "args": {
      "External id": 3784,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2031, "finished": true, "Ev Idx": 3783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564843352.804, "dur": 1.402,
    "args": {
      "External id": 3785,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2032, "finished": true, "Ev Idx": 3784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843353.525, "dur": 0.287,
    "args": {
      "External id": 3786,"Record function id": 0, "finished": true, "Ev Idx": 3785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843354.528, "dur": 3.325,
    "args": {
      "External id": 3787,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2033, "finished": true, "Ev Idx": 3786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564843355.395, "dur": 2.347,
    "args": {
      "External id": 3788,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2033, "finished": true, "Ev Idx": 3787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564843358.555, "dur": 33.546,
    "args": {
      "External id": 3789,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2034, "finished": true, "Ev Idx": 3788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843373.715, "dur": 1.068,
    "args": {
      "External id": 3790,"Record function id": 0, "finished": true, "Ev Idx": 3789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564843392.650, "dur": 1.077,
    "args": {
      "External id": 3791,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2035, "finished": true, "Ev Idx": 3790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564843397.087, "dur": 9.169,
    "args": {
      "External id": 3792,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2036, "finished": true, "Ev Idx": 3791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843409.196, "dur": 1.750,
    "args": {
      "External id": 3793,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2037, "finished": true, "Ev Idx": 3792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843409.950, "dur": 0.402,
    "args": {
      "External id": 3794,"Record function id": 0, "finished": true, "Ev Idx": 3793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843411.623, "dur": 1.306,
    "args": {
      "External id": 3795,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2037, "finished": true, "Ev Idx": 3794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843412.227, "dur": 0.295,
    "args": {
      "External id": 3796,"Record function id": 0, "finished": true, "Ev Idx": 3795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564843413.622, "dur": 1.008,
    "args": {
      "External id": 3797,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2037, "finished": true, "Ev Idx": 3796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843414.075, "dur": 0.255,
    "args": {
      "External id": 3798,"Record function id": 0, "finished": true, "Ev Idx": 3797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564843416.234, "dur": 8.263,
    "args": {
      "External id": 3799,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2037, "finished": true, "Ev Idx": 3798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564843429.596, "dur": 14.284,
    "args": {
      "External id": 3800,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2038, "finished": true, "Ev Idx": 3799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564843430.489, "dur": 0.333,
    "args": {
      "External id": 3801,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2038, "finished": true, "Ev Idx": 3800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564843431.724, "dur": 11.875,
    "args": {
      "External id": 3802,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2038, "finished": true, "Ev Idx": 3801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564843446.797, "dur": 0.201,
    "args": {
      "External id": 3803,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2039, "finished": true, "Ev Idx": 3802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564843457.692, "dur": 0.440,
    "args": {
      "External id": 3804,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2039, "finished": true, "Ev Idx": 3803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564843459.864, "dur": 39.092,
    "args": {
      "External id": 3805,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2039, "finished": true, "Ev Idx": 3804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564843460.569, "dur": 2.441,
    "args": {
      "External id": 3806,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2039, "finished": true, "Ev Idx": 3805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843461.994, "dur": 0.512,
    "args": {
      "External id": 3807,"Record function id": 0, "finished": true, "Ev Idx": 3806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843463.389, "dur": 2.486,
    "args": {
      "External id": 3808,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2040, "finished": true, "Ev Idx": 3807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564843463.740, "dur": 2.010,
    "args": {
      "External id": 3809,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2040, "finished": true, "Ev Idx": 3808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564843466.339, "dur": 1.253,
    "args": {
      "External id": 3810,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2041, "finished": true, "Ev Idx": 3809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843466.965, "dur": 0.269,
    "args": {
      "External id": 3811,"Record function id": 0, "finished": true, "Ev Idx": 3810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843467.871, "dur": 1.343,
    "args": {
      "External id": 3812,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2042, "finished": true, "Ev Idx": 3811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564843468.285, "dur": 0.830,
    "args": {
      "External id": 3813,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2042, "finished": true, "Ev Idx": 3812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564843469.612, "dur": 27.411,
    "args": {
      "External id": 3814,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2043, "finished": true, "Ev Idx": 3813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843480.806, "dur": 1.107,
    "args": {
      "External id": 3815,"Record function id": 0, "finished": true, "Ev Idx": 3814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564843497.624, "dur": 1.044,
    "args": {
      "External id": 3816,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2044, "finished": true, "Ev Idx": 3815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564843500.887, "dur": 1.846,
    "args": {
      "External id": 3817,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2045, "finished": true, "Ev Idx": 3816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843501.848, "dur": 0.389,
    "args": {
      "External id": 3818,"Record function id": 0, "finished": true, "Ev Idx": 3817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564843505.317, "dur": 11.421,
    "args": {
      "External id": 3819,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2046, "finished": true, "Ev Idx": 3818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564843505.892, "dur": 10.712,
    "args": {
      "External id": 3820,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2046, "finished": true, "Ev Idx": 3819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564843506.854, "dur": 1.917,
    "args": {
      "External id": 3821,"Record function id": 0, "finished": true, "Ev Idx": 3820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843507.230, "dur": 1.398,
    "args": {
      "External id": 3822,"Record function id": 0, "finished": true, "Ev Idx": 3821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564843509.073, "dur": 7.094,
    "args": {
      "External id": 3823,"Record function id": 0, "finished": true, "Ev Idx": 3822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843522.355, "dur": 1.619,
    "args": {
      "External id": 3824,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2047, "finished": true, "Ev Idx": 3823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564843522.682, "dur": 1.180,
    "args": {
      "External id": 3825,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2047, "finished": true, "Ev Idx": 3824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564843538.353, "dur": 35.929,
    "args": {
      "External id": 3826,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2048, "finished": true, "Ev Idx": 3825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564843538.813, "dur": 3.098,
    "args": {
      "External id": 3827,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2048, "finished": true, "Ev Idx": 3826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564843540.131, "dur": 1.141,
    "args": {
      "External id": 3828,"Record function id": 0, "finished": true, "Ev Idx": 3827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843540.767, "dur": 0.399,
    "args": {
      "External id": 3829,"Record function id": 0, "finished": true, "Ev Idx": 3828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564843542.424, "dur": 31.717,
    "args": {
      "External id": 3830,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2049, "finished": true, "Ev Idx": 3829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843543.024, "dur": 1.655,
    "args": {
      "External id": 3831,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2049, "finished": true, "Ev Idx": 3830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564843543.532, "dur": 1.045,
    "args": {
      "External id": 3832,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2049, "finished": true, "Ev Idx": 3831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564843545.084, "dur": 26.982,
    "args": {
      "External id": 3833,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2050, "finished": true, "Ev Idx": 3832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843557.165, "dur": 1.086,
    "args": {
      "External id": 3834,"Record function id": 0, "finished": true, "Ev Idx": 3833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564843572.640, "dur": 1.153,
    "args": {
      "External id": 3835,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2051, "finished": true, "Ev Idx": 3834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564843578.991, "dur": 65.302,
    "args": {
      "External id": 3836,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2052, "finished": true, "Ev Idx": 3835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564843656.919, "dur": 0.246,
    "args": {
      "External id": 3837,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2053, "finished": true, "Ev Idx": 3836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564843659.503, "dur": 12.897,
    "args": {
      "External id": 3838,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2053, "finished": true, "Ev Idx": 3837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564843661.333, "dur": 0.395,
    "args": {
      "External id": 3839,"Record function id": 0, "finished": true, "Ev Idx": 3838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564843662.111, "dur": 0.285,
    "args": {
      "External id": 3840,"Record function id": 0, "finished": true, "Ev Idx": 3839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564843675.561, "dur": 15.059,
    "args": {
      "External id": 3841,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2054, "finished": true, "Ev Idx": 3840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564843694.357, "dur": 8.401,
    "args": {
      "External id": 3842,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2055, "finished": true, "Ev Idx": 3841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564843704.926, "dur": 10.485,
    "args": {
      "External id": 3843,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2056, "finished": true, "Ev Idx": 3842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564843717.276, "dur": 9.253,
    "args": {
      "External id": 3844,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2057, "finished": true, "Ev Idx": 3843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564843729.155, "dur": 0.107,
    "args": {
      "External id": 3845,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2058, "finished": true, "Ev Idx": 3844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564843730.247, "dur": 9.575,
    "args": {
      "External id": 3846,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2058, "finished": true, "Ev Idx": 3845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564843754.596, "dur": 54.282,
    "args": {
      "External id": 3847,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2059, "finished": true, "Ev Idx": 3846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564843755.276, "dur": 4.356,
    "args": {
      "External id": 3848,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2059, "finished": true, "Ev Idx": 3847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564843757.010, "dur": 1.853,
    "args": {
      "External id": 3849,"Record function id": 0, "finished": true, "Ev Idx": 3848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843757.795, "dur": 0.840,
    "args": {
      "External id": 3850,"Record function id": 0, "finished": true, "Ev Idx": 3849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564843760.006, "dur": 48.730,
    "args": {
      "External id": 3851,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2060, "finished": true, "Ev Idx": 3850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843760.705, "dur": 2.205,
    "args": {
      "External id": 3852,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2060, "finished": true, "Ev Idx": 3851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564843761.132, "dur": 1.672,
    "args": {
      "External id": 3853,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2060, "finished": true, "Ev Idx": 3852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564843763.294, "dur": 43.247,
    "args": {
      "External id": 3854,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2061, "finished": true, "Ev Idx": 3853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843777.368, "dur": 1.245,
    "args": {
      "External id": 3855,"Record function id": 0, "finished": true, "Ev Idx": 3854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564843807.218, "dur": 1.190,
    "args": {
      "External id": 3856,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2062, "finished": true, "Ev Idx": 3855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564843816.929, "dur": 10.999,
    "args": {
      "External id": 3857,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2063, "finished": true, "Ev Idx": 3856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564843835.745, "dur": 85.842,
    "args": {
      "External id": 3858,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2064, "finished": true, "Ev Idx": 3857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564843836.229, "dur": 3.033,
    "args": {
      "External id": 3859,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2064, "finished": true, "Ev Idx": 3858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564843837.485, "dur": 1.169,
    "args": {
      "External id": 3860,"Record function id": 0, "finished": true, "Ev Idx": 3859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843837.987, "dur": 0.555,
    "args": {
      "External id": 3861,"Record function id": 0, "finished": true, "Ev Idx": 3860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564843839.596, "dur": 81.879,
    "args": {
      "External id": 3862,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2065, "finished": true, "Ev Idx": 3861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843840.095, "dur": 3.007,
    "args": {
      "External id": 3863,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2065, "finished": true, "Ev Idx": 3862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564843840.437, "dur": 2.550,
    "args": {
      "External id": 3864,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2065, "finished": true, "Ev Idx": 3863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564843843.459, "dur": 75.821,
    "args": {
      "External id": 3865,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2066, "finished": true, "Ev Idx": 3864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843894.815, "dur": 1.308,
    "args": {
      "External id": 3866,"Record function id": 0, "finished": true, "Ev Idx": 3865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564843919.924, "dur": 1.210,
    "args": {
      "External id": 3867,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2067, "finished": true, "Ev Idx": 3866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564843923.841, "dur": 9.396,
    "args": {
      "External id": 3868,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2068, "finished": true, "Ev Idx": 3867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564843940.245, "dur": 43.970,
    "args": {
      "External id": 3869,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2069, "finished": true, "Ev Idx": 3868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564843940.878, "dur": 2.920,
    "args": {
      "External id": 3870,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2069, "finished": true, "Ev Idx": 3869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564843941.994, "dur": 1.167,
    "args": {
      "External id": 3871,"Record function id": 0, "finished": true, "Ev Idx": 3870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564843942.562, "dur": 0.466,
    "args": {
      "External id": 3872,"Record function id": 0, "finished": true, "Ev Idx": 3871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564843944.305, "dur": 39.786,
    "args": {
      "External id": 3873,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2070, "finished": true, "Ev Idx": 3872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564843944.919, "dur": 1.652,
    "args": {
      "External id": 3874,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2070, "finished": true, "Ev Idx": 3873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564843945.234, "dur": 1.251,
    "args": {
      "External id": 3875,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2070, "finished": true, "Ev Idx": 3874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564843946.893, "dur": 35.211,
    "args": {
      "External id": 3876,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2071, "finished": true, "Ev Idx": 3875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564843959.420, "dur": 1.079,
    "args": {
      "External id": 3877,"Record function id": 0, "finished": true, "Ev Idx": 3876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564843982.669, "dur": 1.105,
    "args": {
      "External id": 3878,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2072, "finished": true, "Ev Idx": 3877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564843986.871, "dur": 8.335,
    "args": {
      "External id": 3879,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2073, "finished": true, "Ev Idx": 3878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564844015.227, "dur": 0.245,
    "args": {
      "External id": 3880,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2074, "finished": true, "Ev Idx": 3879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564844017.298, "dur": 9.310,
    "args": {
      "External id": 3881,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2074, "finished": true, "Ev Idx": 3880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564844018.676, "dur": 0.367,
    "args": {
      "External id": 3882,"Record function id": 0, "finished": true, "Ev Idx": 3881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564844019.355, "dur": 0.086,
    "args": {
      "External id": 3883,"Record function id": 0, "finished": true, "Ev Idx": 3882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564844029.118, "dur": 12.436,
    "args": {
      "External id": 3884,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2075, "finished": true, "Ev Idx": 3883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564844045.094, "dur": 8.610,
    "args": {
      "External id": 3885,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2076, "finished": true, "Ev Idx": 3884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564844055.397, "dur": 9.228,
    "args": {
      "External id": 3886,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2077, "finished": true, "Ev Idx": 3885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564844066.085, "dur": 9.572,
    "args": {
      "External id": 3887,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2078, "finished": true, "Ev Idx": 3886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564844077.982, "dur": 0.119,
    "args": {
      "External id": 3888,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2079, "finished": true, "Ev Idx": 3887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564844079.008, "dur": 56.241,
    "args": {
      "External id": 3889,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2079, "finished": true, "Ev Idx": 3888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564844158.361, "dur": 47.390,
    "args": {
      "External id": 3890,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2080, "finished": true, "Ev Idx": 3889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564844159.407, "dur": 4.992,
    "args": {
      "External id": 3891,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2080, "finished": true, "Ev Idx": 3890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564844161.017, "dur": 2.292,
    "args": {
      "External id": 3892,"Record function id": 0, "finished": true, "Ev Idx": 3891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844162.202, "dur": 0.868,
    "args": {
      "External id": 3893,"Record function id": 0, "finished": true, "Ev Idx": 3892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564844164.803, "dur": 40.823,
    "args": {
      "External id": 3894,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2081, "finished": true, "Ev Idx": 3893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844165.784, "dur": 3.794,
    "args": {
      "External id": 3895,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2081, "finished": true, "Ev Idx": 3894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844166.229, "dur": 3.124,
    "args": {
      "External id": 3896,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2081, "finished": true, "Ev Idx": 3895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564844169.945, "dur": 33.152,
    "args": {
      "External id": 3897,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2082, "finished": true, "Ev Idx": 3896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564844183.322, "dur": 1.616,
    "args": {
      "External id": 3898,"Record function id": 0, "finished": true, "Ev Idx": 3897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564844203.858, "dur": 1.331,
    "args": {
      "External id": 3899,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2083, "finished": true, "Ev Idx": 3898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844209.205, "dur": 1.714,
    "args": {
      "External id": 3900,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2084, "finished": true, "Ev Idx": 3899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564844212.812, "dur": 2.774,
    "args": {
      "External id": 3901,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2085, "finished": true, "Ev Idx": 3900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844214.051, "dur": 0.591,
    "args": {
      "External id": 3902,"Record function id": 0, "finished": true, "Ev Idx": 3901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564844222.494, "dur": 42.927,
    "args": {
      "External id": 3903,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2086, "finished": true, "Ev Idx": 3902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564844222.848, "dur": 2.372,
    "args": {
      "External id": 3904,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2086, "finished": true, "Ev Idx": 3903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564844223.815, "dur": 0.917,
    "args": {
      "External id": 3905,"Record function id": 0, "finished": true, "Ev Idx": 3904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844224.320, "dur": 0.297,
    "args": {
      "External id": 3906,"Record function id": 0, "finished": true, "Ev Idx": 3905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564844225.561, "dur": 39.764,
    "args": {
      "External id": 3907,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2087, "finished": true, "Ev Idx": 3906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844226.101, "dur": 1.609,
    "args": {
      "External id": 3908,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2087, "finished": true, "Ev Idx": 3907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844226.543, "dur": 1.051,
    "args": {
      "External id": 3909,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2087, "finished": true, "Ev Idx": 3908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564844228.272, "dur": 35.120,
    "args": {
      "External id": 3910,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2088, "finished": true, "Ev Idx": 3909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564844238.488, "dur": 1.086,
    "args": {
      "External id": 3911,"Record function id": 0, "finished": true, "Ev Idx": 3910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564844264.026, "dur": 1.002,
    "args": {
      "External id": 3912,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2089, "finished": true, "Ev Idx": 3911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844267.707, "dur": 1.161,
    "args": {
      "External id": 3913,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2090, "finished": true, "Ev Idx": 3912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564844270.307, "dur": 1.606,
    "args": {
      "External id": 3914,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2091, "finished": true, "Ev Idx": 3913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844271.039, "dur": 0.400,
    "args": {
      "External id": 3915,"Record function id": 0, "finished": true, "Ev Idx": 3914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564844280.049, "dur": 38.325,
    "args": {
      "External id": 3916,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2092, "finished": true, "Ev Idx": 3915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564844280.450, "dur": 3.354,
    "args": {
      "External id": 3917,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2092, "finished": true, "Ev Idx": 3916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564844282.269, "dur": 1.025,
    "args": {
      "External id": 3918,"Record function id": 0, "finished": true, "Ev Idx": 3917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844282.881, "dur": 0.305,
    "args": {
      "External id": 3919,"Record function id": 0, "finished": true, "Ev Idx": 3918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564844284.087, "dur": 34.209,
    "args": {
      "External id": 3920,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2093, "finished": true, "Ev Idx": 3919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844284.497, "dur": 1.617,
    "args": {
      "External id": 3921,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2093, "finished": true, "Ev Idx": 3920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844284.766, "dur": 1.246,
    "args": {
      "External id": 3922,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2093, "finished": true, "Ev Idx": 3921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564844286.434, "dur": 30.020,
    "args": {
      "External id": 3923,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2094, "finished": true, "Ev Idx": 3922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564844295.546, "dur": 1.223,
    "args": {
      "External id": 3924,"Record function id": 0, "finished": true, "Ev Idx": 3923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564844316.981, "dur": 1.021,
    "args": {
      "External id": 3925,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2095, "finished": true, "Ev Idx": 3924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844320.743, "dur": 1.387,
    "args": {
      "External id": 3926,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2096, "finished": true, "Ev Idx": 3925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564844323.443, "dur": 1.857,
    "args": {
      "External id": 3927,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2097, "finished": true, "Ev Idx": 3926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844324.229, "dur": 0.444,
    "args": {
      "External id": 3928,"Record function id": 0, "finished": true, "Ev Idx": 3927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564844328.288, "dur": 2.095,
    "args": {
      "External id": 3929,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2098, "finished": true, "Ev Idx": 3928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844329.646, "dur": 0.259,
    "args": {
      "External id": 3930,"Record function id": 0, "finished": true, "Ev Idx": 3929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564844331.550, "dur": 1.281,
    "args": {
      "External id": 3931,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2098, "finished": true, "Ev Idx": 3930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844332.027, "dur": 0.451,
    "args": {
      "External id": 3932,"Record function id": 0, "finished": true, "Ev Idx": 3931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564844333.820, "dur": 9.754,
    "args": {
      "External id": 3933,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2098, "finished": true, "Ev Idx": 3932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844349.448, "dur": 5.774,
    "args": {
      "External id": 3934,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2099, "finished": true, "Ev Idx": 3933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844352.833, "dur": 0.807,
    "args": {
      "External id": 3935,"Record function id": 0, "finished": true, "Ev Idx": 3934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844357.576, "dur": 3.611,
    "args": {
      "External id": 3936,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2100, "finished": true, "Ev Idx": 3935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844359.531, "dur": 0.611,
    "args": {
      "External id": 3937,"Record function id": 0, "finished": true, "Ev Idx": 3936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564844363.079, "dur": 10.529,
    "args": {
      "External id": 3938,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2101, "finished": true, "Ev Idx": 3937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564844377.644, "dur": 16.756,
    "args": {
      "External id": 3939,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2102, "finished": true, "Ev Idx": 3938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564844397.553, "dur": 9.785,
    "args": {
      "External id": 3940,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2103, "finished": true, "Ev Idx": 3939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564844409.149, "dur": 10.786,
    "args": {
      "External id": 3941,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2104, "finished": true, "Ev Idx": 3940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564844422.039, "dur": 7.823,
    "args": {
      "External id": 3942,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2105, "finished": true, "Ev Idx": 3941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844432.132, "dur": 2.809,
    "args": {
      "External id": 3943,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2106, "finished": true, "Ev Idx": 3942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844433.633, "dur": 0.639,
    "args": {
      "External id": 3944,"Record function id": 0, "finished": true, "Ev Idx": 3943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844436.497, "dur": 2.384,
    "args": {
      "External id": 3945,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2107, "finished": true, "Ev Idx": 3944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844437.960, "dur": 0.381,
    "args": {
      "External id": 3946,"Record function id": 0, "finished": true, "Ev Idx": 3945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564844439.948, "dur": 6.701,
    "args": {
      "External id": 3947,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2108, "finished": true, "Ev Idx": 3946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564844448.475, "dur": 13.372,
    "args": {
      "External id": 3948,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2109, "finished": true, "Ev Idx": 3947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564844464.151, "dur": 6.736,
    "args": {
      "External id": 3949,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2110, "finished": true, "Ev Idx": 3948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564844472.617, "dur": 6.838,
    "args": {
      "External id": 3950,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2111, "finished": true, "Ev Idx": 3949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844503.113, "dur": 3.273,
    "args": {
      "External id": 3951,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2112, "finished": true, "Ev Idx": 3950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844505.136, "dur": 0.517,
    "args": {
      "External id": 3952,"Record function id": 0, "finished": true, "Ev Idx": 3951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844507.054, "dur": 2.541,
    "args": {
      "External id": 3953,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2113, "finished": true, "Ev Idx": 3952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844508.536, "dur": 0.523,
    "args": {
      "External id": 3954,"Record function id": 0, "finished": true, "Ev Idx": 3953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564844510.344, "dur": 2.609,
    "args": {
      "External id": 3955,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2114, "finished": true, "Ev Idx": 3954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844511.424, "dur": 0.917,
    "args": {
      "External id": 3956,"Record function id": 0, "finished": true, "Ev Idx": 3955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844513.774, "dur": 2.081,
    "args": {
      "External id": 3957,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2115, "finished": true, "Ev Idx": 3956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844514.767, "dur": 0.569,
    "args": {
      "External id": 3958,"Record function id": 0, "finished": true, "Ev Idx": 3957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844516.476, "dur": 1.793,
    "args": {
      "External id": 3959,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2116, "finished": true, "Ev Idx": 3958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844517.308, "dur": 0.468,
    "args": {
      "External id": 3960,"Record function id": 0, "finished": true, "Ev Idx": 3959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564844521.146, "dur": 2.864,
    "args": {
      "External id": 3961,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2117, "finished": true, "Ev Idx": 3960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844522.848, "dur": 0.414,
    "args": {
      "External id": 3962,"Record function id": 0, "finished": true, "Ev Idx": 3961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844526.478, "dur": 22.295,
    "args": {
      "External id": 3963,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2118, "finished": true, "Ev Idx": 3962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564844529.272, "dur": 17.581,
    "args": {
      "External id": 3964,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2118, "finished": true, "Ev Idx": 3963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564844531.362, "dur": 3.095,
    "args": {
      "External id": 3965,"Record function id": 0, "finished": true, "Ev Idx": 3964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564844532.454, "dur": 1.738,
    "args": {
      "External id": 3966,"Record function id": 0, "finished": true, "Ev Idx": 3965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564844535.170, "dur": 11.091,
    "args": {
      "External id": 3967,"Record function id": 0, "finished": true, "Ev Idx": 3966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564844547.363, "dur": 1.157,
    "args": {
      "External id": 3968,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2119, "finished": true, "Ev Idx": 3967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844552.179, "dur": 2.307,
    "args": {
      "External id": 3969,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2120, "finished": true, "Ev Idx": 3968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844553.486, "dur": 0.406,
    "args": {
      "External id": 3970,"Record function id": 0, "finished": true, "Ev Idx": 3969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844555.274, "dur": 1.448,
    "args": {
      "External id": 3971,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2121, "finished": true, "Ev Idx": 3970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844555.997, "dur": 0.283,
    "args": {
      "External id": 3972,"Record function id": 0, "finished": true, "Ev Idx": 3971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564844557.233, "dur": 2.736,
    "args": {
      "External id": 3973,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2122, "finished": true, "Ev Idx": 3972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844558.906, "dur": 0.638,
    "args": {
      "External id": 3974,"Record function id": 0, "finished": true, "Ev Idx": 3973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844560.553, "dur": 3.129,
    "args": {
      "External id": 3975,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2123, "finished": true, "Ev Idx": 3974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844562.618, "dur": 0.603,
    "args": {
      "External id": 3976,"Record function id": 0, "finished": true, "Ev Idx": 3975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844564.267, "dur": 2.183,
    "args": {
      "External id": 3977,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2124, "finished": true, "Ev Idx": 3976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844565.711, "dur": 0.293,
    "args": {
      "External id": 3978,"Record function id": 0, "finished": true, "Ev Idx": 3977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564844568.052, "dur": 2.324,
    "args": {
      "External id": 3979,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2125, "finished": true, "Ev Idx": 3978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844569.623, "dur": 0.314,
    "args": {
      "External id": 3980,"Record function id": 0, "finished": true, "Ev Idx": 3979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844571.851, "dur": 11.820,
    "args": {
      "External id": 3981,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2126, "finished": true, "Ev Idx": 3980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564844572.817, "dur": 9.345,
    "args": {
      "External id": 3982,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2126, "finished": true, "Ev Idx": 3981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564844573.618, "dur": 1.705,
    "args": {
      "External id": 3983,"Record function id": 0, "finished": true, "Ev Idx": 3982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564844573.963, "dur": 1.225,
    "args": {
      "External id": 3984,"Record function id": 0, "finished": true, "Ev Idx": 3983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564844575.678, "dur": 6.113,
    "args": {
      "External id": 3985,"Record function id": 0, "finished": true, "Ev Idx": 3984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564844582.572, "dur": 0.912,
    "args": {
      "External id": 3986,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2127, "finished": true, "Ev Idx": 3985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564844585.794, "dur": 1.810,
    "args": {
      "External id": 3987,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2128, "finished": true, "Ev Idx": 3986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844586.723, "dur": 0.417,
    "args": {
      "External id": 3988,"Record function id": 0, "finished": true, "Ev Idx": 3987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564844589.085, "dur": 44.781,
    "args": {
      "External id": 3989,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2129, "finished": true, "Ev Idx": 3988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564844589.898, "dur": 2.831,
    "args": {
      "External id": 3990,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2129, "finished": true, "Ev Idx": 3989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844591.895, "dur": 0.423,
    "args": {
      "External id": 3991,"Record function id": 0, "finished": true, "Ev Idx": 3990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844593.239, "dur": 4.990,
    "args": {
      "External id": 3992,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2130, "finished": true, "Ev Idx": 3991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564844594.969, "dur": 3.096,
    "args": {
      "External id": 3993,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2130, "finished": true, "Ev Idx": 3992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564844598.703, "dur": 1.359,
    "args": {
      "External id": 3994,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2131, "finished": true, "Ev Idx": 3993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844599.424, "dur": 0.281,
    "args": {
      "External id": 3995,"Record function id": 0, "finished": true, "Ev Idx": 3994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844600.362, "dur": 1.730,
    "args": {
      "External id": 3996,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2132, "finished": true, "Ev Idx": 3995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564844601.147, "dur": 0.844,
    "args": {
      "External id": 3997,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2132, "finished": true, "Ev Idx": 3996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564844602.753, "dur": 29.020,
    "args": {
      "External id": 3998,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2133, "finished": true, "Ev Idx": 3997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564844614.354, "dur": 0.930,
    "args": {
      "External id": 3999,"Record function id": 0, "finished": true, "Ev Idx": 3998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564844632.416, "dur": 1.120,
    "args": {
      "External id": 4000,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2134, "finished": true, "Ev Idx": 3999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564844637.122, "dur": 10.484,
    "args": {
      "External id": 4001,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2135, "finished": true, "Ev Idx": 4000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844650.539, "dur": 1.839,
    "args": {
      "External id": 4002,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2136, "finished": true, "Ev Idx": 4001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844651.345, "dur": 0.385,
    "args": {
      "External id": 4003,"Record function id": 0, "finished": true, "Ev Idx": 4002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844653.027, "dur": 1.267,
    "args": {
      "External id": 4004,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2136, "finished": true, "Ev Idx": 4003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844653.608, "dur": 0.296,
    "args": {
      "External id": 4005,"Record function id": 0, "finished": true, "Ev Idx": 4004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564844655.111, "dur": 1.346,
    "args": {
      "External id": 4006,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2136, "finished": true, "Ev Idx": 4005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844655.867, "dur": 0.257,
    "args": {
      "External id": 4007,"Record function id": 0, "finished": true, "Ev Idx": 4006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564844658.149, "dur": 7.220,
    "args": {
      "External id": 4008,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2136, "finished": true, "Ev Idx": 4007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564844670.317, "dur": 13.254,
    "args": {
      "External id": 4009,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2137, "finished": true, "Ev Idx": 4008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564844671.254, "dur": 0.409,
    "args": {
      "External id": 4010,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2137, "finished": true, "Ev Idx": 4009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564844672.459, "dur": 10.798,
    "args": {
      "External id": 4011,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2137, "finished": true, "Ev Idx": 4010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564844686.564, "dur": 0.187,
    "args": {
      "External id": 4012,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2138, "finished": true, "Ev Idx": 4011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564844696.979, "dur": 0.473,
    "args": {
      "External id": 4013,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2138, "finished": true, "Ev Idx": 4012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564844699.117, "dur": 38.772,
    "args": {
      "External id": 4014,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2138, "finished": true, "Ev Idx": 4013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564844699.583, "dur": 1.999,
    "args": {
      "External id": 4015,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2138, "finished": true, "Ev Idx": 4014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844700.652, "dur": 0.387,
    "args": {
      "External id": 4016,"Record function id": 0, "finished": true, "Ev Idx": 4015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844701.952, "dur": 2.400,
    "args": {
      "External id": 4017,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2139, "finished": true, "Ev Idx": 4016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844702.271, "dur": 1.948,
    "args": {
      "External id": 4018,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2139, "finished": true, "Ev Idx": 4017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564844704.784, "dur": 2.514,
    "args": {
      "External id": 4019,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2140, "finished": true, "Ev Idx": 4018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844706.706, "dur": 0.219,
    "args": {
      "External id": 4020,"Record function id": 0, "finished": true, "Ev Idx": 4019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844707.703, "dur": 2.533,
    "args": {
      "External id": 4021,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2141, "finished": true, "Ev Idx": 4020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844708.084, "dur": 2.090,
    "args": {
      "External id": 4022,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2141, "finished": true, "Ev Idx": 4021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564844710.694, "dur": 25.232,
    "args": {
      "External id": 4023,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2142, "finished": true, "Ev Idx": 4022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564844719.271, "dur": 1.075,
    "args": {
      "External id": 4024,"Record function id": 0, "finished": true, "Ev Idx": 4023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564844736.505, "dur": 1.106,
    "args": {
      "External id": 4025,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2143, "finished": true, "Ev Idx": 4024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564844739.691, "dur": 1.796,
    "args": {
      "External id": 4026,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2144, "finished": true, "Ev Idx": 4025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844740.613, "dur": 0.388,
    "args": {
      "External id": 4027,"Record function id": 0, "finished": true, "Ev Idx": 4026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564844744.192, "dur": 11.522,
    "args": {
      "External id": 4028,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2145, "finished": true, "Ev Idx": 4027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564844744.762, "dur": 10.815,
    "args": {
      "External id": 4029,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2145, "finished": true, "Ev Idx": 4028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564844745.729, "dur": 2.065,
    "args": {
      "External id": 4030,"Record function id": 0, "finished": true, "Ev Idx": 4029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564844746.126, "dur": 1.519,
    "args": {
      "External id": 4031,"Record function id": 0, "finished": true, "Ev Idx": 4030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564844748.081, "dur": 7.023,
    "args": {
      "External id": 4032,"Record function id": 0, "finished": true, "Ev Idx": 4031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844761.236, "dur": 1.754,
    "args": {
      "External id": 4033,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2146, "finished": true, "Ev Idx": 4032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844761.563, "dur": 1.301,
    "args": {
      "External id": 4034,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2146, "finished": true, "Ev Idx": 4033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564844775.153, "dur": 33.318,
    "args": {
      "External id": 4035,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2147, "finished": true, "Ev Idx": 4034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564844775.918, "dur": 2.885,
    "args": {
      "External id": 4036,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2147, "finished": true, "Ev Idx": 4035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564844777.243, "dur": 1.010,
    "args": {
      "External id": 4037,"Record function id": 0, "finished": true, "Ev Idx": 4036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844777.794, "dur": 0.344,
    "args": {
      "External id": 4038,"Record function id": 0, "finished": true, "Ev Idx": 4037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564844779.279, "dur": 29.056,
    "args": {
      "External id": 4039,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2148, "finished": true, "Ev Idx": 4038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844779.834, "dur": 2.174,
    "args": {
      "External id": 4040,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2148, "finished": true, "Ev Idx": 4039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844780.167, "dur": 1.731,
    "args": {
      "External id": 4041,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2148, "finished": true, "Ev Idx": 4040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564844782.403, "dur": 24.047,
    "args": {
      "External id": 4042,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2149, "finished": true, "Ev Idx": 4041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564844791.982, "dur": 1.074,
    "args": {
      "External id": 4043,"Record function id": 0, "finished": true, "Ev Idx": 4042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564844807.048, "dur": 0.958,
    "args": {
      "External id": 4044,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2150, "finished": true, "Ev Idx": 4043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564844813.007, "dur": 9.699,
    "args": {
      "External id": 4045,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2151, "finished": true, "Ev Idx": 4044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564844831.786, "dur": 0.332,
    "args": {
      "External id": 4046,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2152, "finished": true, "Ev Idx": 4045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564844834.217, "dur": 12.008,
    "args": {
      "External id": 4047,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2152, "finished": true, "Ev Idx": 4046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564844835.922, "dur": 0.330,
    "args": {
      "External id": 4048,"Record function id": 0, "finished": true, "Ev Idx": 4047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564844836.673, "dur": 0.263,
    "args": {
      "External id": 4049,"Record function id": 0, "finished": true, "Ev Idx": 4048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564844848.907, "dur": 13.135,
    "args": {
      "External id": 4050,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2153, "finished": true, "Ev Idx": 4049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564844865.676, "dur": 8.913,
    "args": {
      "External id": 4051,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2154, "finished": true, "Ev Idx": 4050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564844876.635, "dur": 10.436,
    "args": {
      "External id": 4052,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2155, "finished": true, "Ev Idx": 4051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564844888.726, "dur": 7.782,
    "args": {
      "External id": 4053,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2156, "finished": true, "Ev Idx": 4052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564844899.021, "dur": 0.155,
    "args": {
      "External id": 4054,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2157, "finished": true, "Ev Idx": 4053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564844900.108, "dur": 9.780,
    "args": {
      "External id": 4055,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2157, "finished": true, "Ev Idx": 4054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564844923.373, "dur": 106.385,
    "args": {
      "External id": 4056,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2158, "finished": true, "Ev Idx": 4055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564844923.773, "dur": 2.755,
    "args": {
      "External id": 4057,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2158, "finished": true, "Ev Idx": 4056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564844924.704, "dur": 1.222,
    "args": {
      "External id": 4058,"Record function id": 0, "finished": true, "Ev Idx": 4057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564844925.382, "dur": 0.413,
    "args": {
      "External id": 4059,"Record function id": 0, "finished": true, "Ev Idx": 4058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564844926.884, "dur": 102.739,
    "args": {
      "External id": 4060,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2159, "finished": true, "Ev Idx": 4059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564844927.340, "dur": 2.265,
    "args": {
      "External id": 4061,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2159, "finished": true, "Ev Idx": 4060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564844927.613, "dur": 1.893,
    "args": {
      "External id": 4062,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2159, "finished": true, "Ev Idx": 4061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564844929.963, "dur": 97.440,
    "args": {
      "External id": 4063,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2160, "finished": true, "Ev Idx": 4062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564844999.370, "dur": 1.319,
    "args": {
      "External id": 4064,"Record function id": 0, "finished": true, "Ev Idx": 4063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564845028.041, "dur": 1.234,
    "args": {
      "External id": 4065,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2161, "finished": true, "Ev Idx": 4064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564845039.405, "dur": 10.995,
    "args": {
      "External id": 4066,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2162, "finished": true, "Ev Idx": 4065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564845058.529, "dur": 78.329,
    "args": {
      "External id": 4067,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2163, "finished": true, "Ev Idx": 4066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564845059.130, "dur": 3.034,
    "args": {
      "External id": 4068,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2163, "finished": true, "Ev Idx": 4067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564845060.257, "dur": 1.344,
    "args": {
      "External id": 4069,"Record function id": 0, "finished": true, "Ev Idx": 4068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845060.882, "dur": 0.598,
    "args": {
      "External id": 4070,"Record function id": 0, "finished": true, "Ev Idx": 4069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564845062.632, "dur": 74.100,
    "args": {
      "External id": 4071,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2164, "finished": true, "Ev Idx": 4070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845063.319, "dur": 2.106,
    "args": {
      "External id": 4072,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2164, "finished": true, "Ev Idx": 4071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564845063.687, "dur": 1.619,
    "args": {
      "External id": 4073,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2164, "finished": true, "Ev Idx": 4072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564845065.873, "dur": 68.323,
    "args": {
      "External id": 4074,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2165, "finished": true, "Ev Idx": 4073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564845110.002, "dur": 1.295,
    "args": {
      "External id": 4075,"Record function id": 0, "finished": true, "Ev Idx": 4074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564845134.928, "dur": 1.456,
    "args": {
      "External id": 4076,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2166, "finished": true, "Ev Idx": 4075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564845139.058, "dur": 9.483,
    "args": {
      "External id": 4077,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2167, "finished": true, "Ev Idx": 4076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564845156.391, "dur": 39.302,
    "args": {
      "External id": 4078,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2168, "finished": true, "Ev Idx": 4077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564845157.183, "dur": 2.806,
    "args": {
      "External id": 4079,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2168, "finished": true, "Ev Idx": 4078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564845158.182, "dur": 1.225,
    "args": {
      "External id": 4080,"Record function id": 0, "finished": true, "Ev Idx": 4079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845158.814, "dur": 0.464,
    "args": {
      "External id": 4081,"Record function id": 0, "finished": true, "Ev Idx": 4080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564845160.497, "dur": 35.117,
    "args": {
      "External id": 4082,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2169, "finished": true, "Ev Idx": 4081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845161.166, "dur": 1.847,
    "args": {
      "External id": 4083,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2169, "finished": true, "Ev Idx": 4082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564845161.499, "dur": 1.396,
    "args": {
      "External id": 4084,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2169, "finished": true, "Ev Idx": 4083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564845163.351, "dur": 30.434,
    "args": {
      "External id": 4085,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2170, "finished": true, "Ev Idx": 4084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564845173.291, "dur": 1.326,
    "args": {
      "External id": 4086,"Record function id": 0, "finished": true, "Ev Idx": 4085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564845194.399, "dur": 0.929,
    "args": {
      "External id": 4087,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2171, "finished": true, "Ev Idx": 4086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564845198.274, "dur": 9.986,
    "args": {
      "External id": 4088,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2172, "finished": true, "Ev Idx": 4087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564845229.852, "dur": 0.192,
    "args": {
      "External id": 4089,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2173, "finished": true, "Ev Idx": 4088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564845231.995, "dur": 11.212,
    "args": {
      "External id": 4090,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2173, "finished": true, "Ev Idx": 4089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564845233.584, "dur": 0.492,
    "args": {
      "External id": 4091,"Record function id": 0, "finished": true, "Ev Idx": 4090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564845234.404, "dur": 0.073,
    "args": {
      "External id": 4092,"Record function id": 0, "finished": true, "Ev Idx": 4091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564845246.224, "dur": 12.411,
    "args": {
      "External id": 4093,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2174, "finished": true, "Ev Idx": 4092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564845262.311, "dur": 8.521,
    "args": {
      "External id": 4094,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2175, "finished": true, "Ev Idx": 4093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564845272.603, "dur": 8.662,
    "args": {
      "External id": 4095,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2176, "finished": true, "Ev Idx": 4094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564845282.931, "dur": 9.964,
    "args": {
      "External id": 4096,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2177, "finished": true, "Ev Idx": 4095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564845328.318, "dur": 0.108,
    "args": {
      "External id": 4097,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2178, "finished": true, "Ev Idx": 4096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564845329.412, "dur": 7.757,
    "args": {
      "External id": 4098,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2178, "finished": true, "Ev Idx": 4097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564845355.677, "dur": 38.260,
    "args": {
      "External id": 4099,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2179, "finished": true, "Ev Idx": 4098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564845356.118, "dur": 4.126,
    "args": {
      "External id": 4100,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2179, "finished": true, "Ev Idx": 4099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564845358.332, "dur": 1.251,
    "args": {
      "External id": 4101,"Record function id": 0, "finished": true, "Ev Idx": 4100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845358.892, "dur": 0.519,
    "args": {
      "External id": 4102,"Record function id": 0, "finished": true, "Ev Idx": 4101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564845360.545, "dur": 33.271,
    "args": {
      "External id": 4103,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2180, "finished": true, "Ev Idx": 4102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845361.029, "dur": 1.536,
    "args": {
      "External id": 4104,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2180, "finished": true, "Ev Idx": 4103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564845361.331, "dur": 1.115,
    "args": {
      "External id": 4105,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2180, "finished": true, "Ev Idx": 4104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564845362.889, "dur": 28.829,
    "args": {
      "External id": 4106,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2181, "finished": true, "Ev Idx": 4105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564845374.132, "dur": 1.100,
    "args": {
      "External id": 4107,"Record function id": 0, "finished": true, "Ev Idx": 4106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564845392.332, "dur": 1.149,
    "args": {
      "External id": 4108,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2182, "finished": true, "Ev Idx": 4107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564845396.953, "dur": 1.348,
    "args": {
      "External id": 4109,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2183, "finished": true, "Ev Idx": 4108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564845400.159, "dur": 2.745,
    "args": {
      "External id": 4110,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2184, "finished": true, "Ev Idx": 4109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845401.314, "dur": 0.695,
    "args": {
      "External id": 4111,"Record function id": 0, "finished": true, "Ev Idx": 4110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564845410.075, "dur": 39.617,
    "args": {
      "External id": 4112,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2185, "finished": true, "Ev Idx": 4111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564845410.379, "dur": 2.349,
    "args": {
      "External id": 4113,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2185, "finished": true, "Ev Idx": 4112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564845411.354, "dur": 0.914,
    "args": {
      "External id": 4114,"Record function id": 0, "finished": true, "Ev Idx": 4113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845411.862, "dur": 0.304,
    "args": {
      "External id": 4115,"Record function id": 0, "finished": true, "Ev Idx": 4114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564845413.020, "dur": 36.574,
    "args": {
      "External id": 4116,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2186, "finished": true, "Ev Idx": 4115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845413.423, "dur": 1.225,
    "args": {
      "External id": 4117,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2186, "finished": true, "Ev Idx": 4116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564845413.680, "dur": 0.872,
    "args": {
      "External id": 4118,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2186, "finished": true, "Ev Idx": 4117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564845414.943, "dur": 32.785,
    "args": {
      "External id": 4119,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2187, "finished": true, "Ev Idx": 4118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564845424.997, "dur": 1.082,
    "args": {
      "External id": 4120,"Record function id": 0, "finished": true, "Ev Idx": 4119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564845448.323, "dur": 0.981,
    "args": {
      "External id": 4121,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2188, "finished": true, "Ev Idx": 4120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564845452.002, "dur": 1.153,
    "args": {
      "External id": 4122,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2189, "finished": true, "Ev Idx": 4121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564845454.705, "dur": 1.887,
    "args": {
      "External id": 4123,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2190, "finished": true, "Ev Idx": 4122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845455.502, "dur": 0.514,
    "args": {
      "External id": 4124,"Record function id": 0, "finished": true, "Ev Idx": 4123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564845488.161, "dur": 32.355,
    "args": {
      "External id": 4125,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2191, "finished": true, "Ev Idx": 4124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564845488.439, "dur": 2.250,
    "args": {
      "External id": 4126,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2191, "finished": true, "Ev Idx": 4125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564845489.371, "dur": 0.877,
    "args": {
      "External id": 4127,"Record function id": 0, "finished": true, "Ev Idx": 4126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845489.833, "dur": 0.290,
    "args": {
      "External id": 4128,"Record function id": 0, "finished": true, "Ev Idx": 4127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564845490.992, "dur": 29.442,
    "args": {
      "External id": 4129,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2192, "finished": true, "Ev Idx": 4128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845491.366, "dur": 1.262,
    "args": {
      "External id": 4130,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2192, "finished": true, "Ev Idx": 4129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564845491.618, "dur": 0.912,
    "args": {
      "External id": 4131,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2192, "finished": true, "Ev Idx": 4130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564845492.939, "dur": 25.709,
    "args": {
      "External id": 4132,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2193, "finished": true, "Ev Idx": 4131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564845499.739, "dur": 1.103,
    "args": {
      "External id": 4133,"Record function id": 0, "finished": true, "Ev Idx": 4132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564845519.204, "dur": 0.928,
    "args": {
      "External id": 4134,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2194, "finished": true, "Ev Idx": 4133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564845523.018, "dur": 1.242,
    "args": {
      "External id": 4135,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2195, "finished": true, "Ev Idx": 4134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564845525.786, "dur": 1.672,
    "args": {
      "External id": 4136,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2196, "finished": true, "Ev Idx": 4135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845526.579, "dur": 0.384,
    "args": {
      "External id": 4137,"Record function id": 0, "finished": true, "Ev Idx": 4136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564845530.521, "dur": 2.064,
    "args": {
      "External id": 4138,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2197, "finished": true, "Ev Idx": 4137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845531.707, "dur": 0.342,
    "args": {
      "External id": 4139,"Record function id": 0, "finished": true, "Ev Idx": 4138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564845533.736, "dur": 1.084,
    "args": {
      "External id": 4140,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2197, "finished": true, "Ev Idx": 4139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845534.208, "dur": 0.284,
    "args": {
      "External id": 4141,"Record function id": 0, "finished": true, "Ev Idx": 4140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564845535.671, "dur": 10.195,
    "args": {
      "External id": 4142,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2197, "finished": true, "Ev Idx": 4141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845551.704, "dur": 4.906,
    "args": {
      "External id": 4143,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2198, "finished": true, "Ev Idx": 4142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845554.346, "dur": 0.600,
    "args": {
      "External id": 4144,"Record function id": 0, "finished": true, "Ev Idx": 4143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845558.811, "dur": 1.763,
    "args": {
      "External id": 4145,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2199, "finished": true, "Ev Idx": 4144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845559.649, "dur": 0.338,
    "args": {
      "External id": 4146,"Record function id": 0, "finished": true, "Ev Idx": 4145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564845562.544, "dur": 9.918,
    "args": {
      "External id": 4147,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2200, "finished": true, "Ev Idx": 4146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564845576.675, "dur": 60.806,
    "args": {
      "External id": 4148,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2201, "finished": true, "Ev Idx": 4147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564845640.813, "dur": 8.212,
    "args": {
      "External id": 4149,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2202, "finished": true, "Ev Idx": 4148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564845651.007, "dur": 8.879,
    "args": {
      "External id": 4150,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2203, "finished": true, "Ev Idx": 4149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564845662.594, "dur": 6.585,
    "args": {
      "External id": 4151,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2204, "finished": true, "Ev Idx": 4150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845672.182, "dur": 2.700,
    "args": {
      "External id": 4152,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2205, "finished": true, "Ev Idx": 4151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845673.423, "dur": 0.558,
    "args": {
      "External id": 4153,"Record function id": 0, "finished": true, "Ev Idx": 4152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845676.503, "dur": 1.709,
    "args": {
      "External id": 4154,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2206, "finished": true, "Ev Idx": 4153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845677.315, "dur": 0.340,
    "args": {
      "External id": 4155,"Record function id": 0, "finished": true, "Ev Idx": 4154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564845679.443, "dur": 7.943,
    "args": {
      "External id": 4156,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2207, "finished": true, "Ev Idx": 4155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564845689.719, "dur": 9.850,
    "args": {
      "External id": 4157,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2208, "finished": true, "Ev Idx": 4156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564845701.509, "dur": 6.731,
    "args": {
      "External id": 4158,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2209, "finished": true, "Ev Idx": 4157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564845710.065, "dur": 6.121,
    "args": {
      "External id": 4159,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2210, "finished": true, "Ev Idx": 4158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845741.112, "dur": 2.758,
    "args": {
      "External id": 4160,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2211, "finished": true, "Ev Idx": 4159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845742.624, "dur": 0.507,
    "args": {
      "External id": 4161,"Record function id": 0, "finished": true, "Ev Idx": 4160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845744.526, "dur": 1.721,
    "args": {
      "External id": 4162,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2212, "finished": true, "Ev Idx": 4161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845745.446, "dur": 0.338,
    "args": {
      "External id": 4163,"Record function id": 0, "finished": true, "Ev Idx": 4162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564845747.095, "dur": 2.665,
    "args": {
      "External id": 4164,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2213, "finished": true, "Ev Idx": 4163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845748.281, "dur": 0.746,
    "args": {
      "External id": 4165,"Record function id": 0, "finished": true, "Ev Idx": 4164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845750.476, "dur": 1.633,
    "args": {
      "External id": 4166,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2214, "finished": true, "Ev Idx": 4165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845751.267, "dur": 0.371,
    "args": {
      "External id": 4167,"Record function id": 0, "finished": true, "Ev Idx": 4166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845752.750, "dur": 1.586,
    "args": {
      "External id": 4168,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2215, "finished": true, "Ev Idx": 4167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845753.434, "dur": 0.445,
    "args": {
      "External id": 4169,"Record function id": 0, "finished": true, "Ev Idx": 4168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564845757.440, "dur": 3.210,
    "args": {
      "External id": 4170,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2216, "finished": true, "Ev Idx": 4169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845759.412, "dur": 0.536,
    "args": {
      "External id": 4171,"Record function id": 0, "finished": true, "Ev Idx": 4170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845763.317, "dur": 21.265,
    "args": {
      "External id": 4172,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2217, "finished": true, "Ev Idx": 4171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564845766.030, "dur": 16.240,
    "args": {
      "External id": 4173,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2217, "finished": true, "Ev Idx": 4172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564845767.954, "dur": 3.186,
    "args": {
      "External id": 4174,"Record function id": 0, "finished": true, "Ev Idx": 4173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564845768.901, "dur": 1.996,
    "args": {
      "External id": 4175,"Record function id": 0, "finished": true, "Ev Idx": 4174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564845771.766, "dur": 9.871,
    "args": {
      "External id": 4176,"Record function id": 0, "finished": true, "Ev Idx": 4175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564845782.772, "dur": 1.509,
    "args": {
      "External id": 4177,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2218, "finished": true, "Ev Idx": 4176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845787.887, "dur": 2.099,
    "args": {
      "External id": 4178,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2219, "finished": true, "Ev Idx": 4177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845788.983, "dur": 0.409,
    "args": {
      "External id": 4179,"Record function id": 0, "finished": true, "Ev Idx": 4178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845790.588, "dur": 1.665,
    "args": {
      "External id": 4180,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2220, "finished": true, "Ev Idx": 4179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845791.508, "dur": 0.281,
    "args": {
      "External id": 4181,"Record function id": 0, "finished": true, "Ev Idx": 4180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564845792.740, "dur": 1.636,
    "args": {
      "External id": 4182,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2221, "finished": true, "Ev Idx": 4181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845793.444, "dur": 0.504,
    "args": {
      "External id": 4183,"Record function id": 0, "finished": true, "Ev Idx": 4182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845794.959, "dur": 1.489,
    "args": {
      "External id": 4184,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2222, "finished": true, "Ev Idx": 4183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845795.634, "dur": 0.307,
    "args": {
      "External id": 4185,"Record function id": 0, "finished": true, "Ev Idx": 4184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845797.139, "dur": 1.404,
    "args": {
      "External id": 4186,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2223, "finished": true, "Ev Idx": 4185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845797.797, "dur": 0.270,
    "args": {
      "External id": 4187,"Record function id": 0, "finished": true, "Ev Idx": 4186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564845800.143, "dur": 1.630,
    "args": {
      "External id": 4188,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2224, "finished": true, "Ev Idx": 4187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845801.022, "dur": 0.319,
    "args": {
      "External id": 4189,"Record function id": 0, "finished": true, "Ev Idx": 4188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845803.325, "dur": 11.235,
    "args": {
      "External id": 4190,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2225, "finished": true, "Ev Idx": 4189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564845804.229, "dur": 8.486,
    "args": {
      "External id": 4191,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2225, "finished": true, "Ev Idx": 4190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564845804.914, "dur": 1.813,
    "args": {
      "External id": 4192,"Record function id": 0, "finished": true, "Ev Idx": 4191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564845805.244, "dur": 1.351,
    "args": {
      "External id": 4193,"Record function id": 0, "finished": true, "Ev Idx": 4192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564845807.071, "dur": 5.332,
    "args": {
      "External id": 4194,"Record function id": 0, "finished": true, "Ev Idx": 4193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564845813.102, "dur": 1.287,
    "args": {
      "External id": 4195,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2226, "finished": true, "Ev Idx": 4194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564845816.842, "dur": 1.915,
    "args": {
      "External id": 4196,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2227, "finished": true, "Ev Idx": 4195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845817.857, "dur": 0.392,
    "args": {
      "External id": 4197,"Record function id": 0, "finished": true, "Ev Idx": 4196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564845820.347, "dur": 48.967,
    "args": {
      "External id": 4198,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2228, "finished": true, "Ev Idx": 4197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564845821.099, "dur": 2.528,
    "args": {
      "External id": 4199,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2228, "finished": true, "Ev Idx": 4198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845822.874, "dur": 0.312,
    "args": {
      "External id": 4200,"Record function id": 0, "finished": true, "Ev Idx": 4199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845823.982, "dur": 4.470,
    "args": {
      "External id": 4201,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2229, "finished": true, "Ev Idx": 4200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564845825.767, "dur": 2.540,
    "args": {
      "External id": 4202,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2229, "finished": true, "Ev Idx": 4201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564845829.124, "dur": 1.370,
    "args": {
      "External id": 4203,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2230, "finished": true, "Ev Idx": 4202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845829.839, "dur": 0.283,
    "args": {
      "External id": 4204,"Record function id": 0, "finished": true, "Ev Idx": 4203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845830.809, "dur": 1.739,
    "args": {
      "External id": 4205,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2231, "finished": true, "Ev Idx": 4204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564845831.651, "dur": 0.791,
    "args": {
      "External id": 4206,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2231, "finished": true, "Ev Idx": 4205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564845833.264, "dur": 33.882,
    "args": {
      "External id": 4207,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2232, "finished": true, "Ev Idx": 4206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564845846.966, "dur": 1.145,
    "args": {
      "External id": 4208,"Record function id": 0, "finished": true, "Ev Idx": 4207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564845867.771, "dur": 1.190,
    "args": {
      "External id": 4209,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2233, "finished": true, "Ev Idx": 4208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564845872.823, "dur": 10.449,
    "args": {
      "External id": 4210,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2234, "finished": true, "Ev Idx": 4209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845886.467, "dur": 1.951,
    "args": {
      "External id": 4211,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2235, "finished": true, "Ev Idx": 4210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845887.255, "dur": 0.422,
    "args": {
      "External id": 4212,"Record function id": 0, "finished": true, "Ev Idx": 4211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845889.048, "dur": 1.385,
    "args": {
      "External id": 4213,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2235, "finished": true, "Ev Idx": 4212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845889.712, "dur": 0.320,
    "args": {
      "External id": 4214,"Record function id": 0, "finished": true, "Ev Idx": 4213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564845891.102, "dur": 1.009,
    "args": {
      "External id": 4215,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2235, "finished": true, "Ev Idx": 4214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845891.561, "dur": 0.220,
    "args": {
      "External id": 4216,"Record function id": 0, "finished": true, "Ev Idx": 4215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564845893.917, "dur": 7.739,
    "args": {
      "External id": 4217,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2235, "finished": true, "Ev Idx": 4216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564845907.167, "dur": 14.424,
    "args": {
      "External id": 4218,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2236, "finished": true, "Ev Idx": 4217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564845908.100, "dur": 0.374,
    "args": {
      "External id": 4219,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2236, "finished": true, "Ev Idx": 4218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564845909.163, "dur": 12.146,
    "args": {
      "External id": 4220,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2236, "finished": true, "Ev Idx": 4219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564845924.535, "dur": 0.177,
    "args": {
      "External id": 4221,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2237, "finished": true, "Ev Idx": 4220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564845935.540, "dur": 0.490,
    "args": {
      "External id": 4222,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2237, "finished": true, "Ev Idx": 4221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564845937.845, "dur": 41.565,
    "args": {
      "External id": 4223,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2237, "finished": true, "Ev Idx": 4222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564845938.459, "dur": 2.182,
    "args": {
      "External id": 4224,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2237, "finished": true, "Ev Idx": 4223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845939.799, "dur": 0.358,
    "args": {
      "External id": 4225,"Record function id": 0, "finished": true, "Ev Idx": 4224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845940.976, "dur": 3.269,
    "args": {
      "External id": 4226,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2238, "finished": true, "Ev Idx": 4225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564845941.324, "dur": 2.790,
    "args": {
      "External id": 4227,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2238, "finished": true, "Ev Idx": 4226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564845944.696, "dur": 1.482,
    "args": {
      "External id": 4228,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2239, "finished": true, "Ev Idx": 4227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845945.545, "dur": 0.265,
    "args": {
      "External id": 4229,"Record function id": 0, "finished": true, "Ev Idx": 4228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564845946.453, "dur": 3.248,
    "args": {
      "External id": 4230,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2240, "finished": true, "Ev Idx": 4229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564845946.690, "dur": 2.928,
    "args": {
      "External id": 4231,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2240, "finished": true, "Ev Idx": 4230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564845950.101, "dur": 27.032,
    "args": {
      "External id": 4232,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2241, "finished": true, "Ev Idx": 4231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564845959.951, "dur": 1.173,
    "args": {
      "External id": 4233,"Record function id": 0, "finished": true, "Ev Idx": 4232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564845977.790, "dur": 1.327,
    "args": {
      "External id": 4234,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2242, "finished": true, "Ev Idx": 4233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564845981.459, "dur": 1.775,
    "args": {
      "External id": 4235,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2243, "finished": true, "Ev Idx": 4234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564845982.344, "dur": 0.385,
    "args": {
      "External id": 4236,"Record function id": 0, "finished": true, "Ev Idx": 4235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564845985.979, "dur": 11.246,
    "args": {
      "External id": 4237,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2244, "finished": true, "Ev Idx": 4236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564845986.722, "dur": 10.360,
    "args": {
      "External id": 4238,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2244, "finished": true, "Ev Idx": 4237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564845987.700, "dur": 1.892,
    "args": {
      "External id": 4239,"Record function id": 0, "finished": true, "Ev Idx": 4238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564845988.039, "dur": 1.398,
    "args": {
      "External id": 4240,"Record function id": 0, "finished": true, "Ev Idx": 4239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564845989.924, "dur": 6.772,
    "args": {
      "External id": 4241,"Record function id": 0, "finished": true, "Ev Idx": 4240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564846002.857, "dur": 1.878,
    "args": {
      "External id": 4242,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2245, "finished": true, "Ev Idx": 4241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846003.147, "dur": 1.487,
    "args": {
      "External id": 4243,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2245, "finished": true, "Ev Idx": 4242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564846018.322, "dur": 33.978,
    "args": {
      "External id": 4244,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2246, "finished": true, "Ev Idx": 4243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564846018.907, "dur": 3.042,
    "args": {
      "External id": 4245,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2246, "finished": true, "Ev Idx": 4244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564846020.322, "dur": 1.016,
    "args": {
      "External id": 4246,"Record function id": 0, "finished": true, "Ev Idx": 4245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846020.722, "dur": 0.512,
    "args": {
      "External id": 4247,"Record function id": 0, "finished": true, "Ev Idx": 4246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564846022.323, "dur": 29.802,
    "args": {
      "External id": 4248,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2247, "finished": true, "Ev Idx": 4247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564846022.964, "dur": 2.126,
    "args": {
      "External id": 4249,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2247, "finished": true, "Ev Idx": 4248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846023.359, "dur": 1.628,
    "args": {
      "External id": 4250,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2247, "finished": true, "Ev Idx": 4249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564846025.524, "dur": 24.734,
    "args": {
      "External id": 4251,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2248, "finished": true, "Ev Idx": 4250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564846035.501, "dur": 1.031,
    "args": {
      "External id": 4252,"Record function id": 0, "finished": true, "Ev Idx": 4251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564846050.838, "dur": 0.977,
    "args": {
      "External id": 4253,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2249, "finished": true, "Ev Idx": 4252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564846057.180, "dur": 9.696,
    "args": {
      "External id": 4254,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2250, "finished": true, "Ev Idx": 4253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564846095.967, "dur": 0.443,
    "args": {
      "External id": 4255,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2251, "finished": true, "Ev Idx": 4254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564846099.093, "dur": 16.014,
    "args": {
      "External id": 4256,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2251, "finished": true, "Ev Idx": 4255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564846101.561, "dur": 0.497,
    "args": {
      "External id": 4257,"Record function id": 0, "finished": true, "Ev Idx": 4256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564846102.760, "dur": 0.397,
    "args": {
      "External id": 4258,"Record function id": 0, "finished": true, "Ev Idx": 4257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564846118.325, "dur": 14.390,
    "args": {
      "External id": 4259,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2252, "finished": true, "Ev Idx": 4258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564846137.312, "dur": 10.210,
    "args": {
      "External id": 4260,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2253, "finished": true, "Ev Idx": 4259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564846149.746, "dur": 10.341,
    "args": {
      "External id": 4261,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2254, "finished": true, "Ev Idx": 4260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564846162.019, "dur": 9.232,
    "args": {
      "External id": 4262,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2255, "finished": true, "Ev Idx": 4261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564846173.979, "dur": 0.092,
    "args": {
      "External id": 4263,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2256, "finished": true, "Ev Idx": 4262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564846175.058, "dur": 10.059,
    "args": {
      "External id": 4264,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2256, "finished": true, "Ev Idx": 4263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564846200.209, "dur": 54.555,
    "args": {
      "External id": 4265,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2257, "finished": true, "Ev Idx": 4264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564846200.707, "dur": 3.603,
    "args": {
      "External id": 4266,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2257, "finished": true, "Ev Idx": 4265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564846202.043, "dur": 1.583,
    "args": {
      "External id": 4267,"Record function id": 0, "finished": true, "Ev Idx": 4266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846202.797, "dur": 0.682,
    "args": {
      "External id": 4268,"Record function id": 0, "finished": true, "Ev Idx": 4267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564846204.676, "dur": 49.941,
    "args": {
      "External id": 4269,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2258, "finished": true, "Ev Idx": 4268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564846205.455, "dur": 3.275,
    "args": {
      "External id": 4270,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2258, "finished": true, "Ev Idx": 4269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846205.819, "dur": 2.790,
    "args": {
      "External id": 4271,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2258, "finished": true, "Ev Idx": 4270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564846209.152, "dur": 43.072,
    "args": {
      "External id": 4272,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2259, "finished": true, "Ev Idx": 4271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564846223.105, "dur": 1.263,
    "args": {
      "External id": 4273,"Record function id": 0, "finished": true, "Ev Idx": 4272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564846252.849, "dur": 1.379,
    "args": {
      "External id": 4274,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2260, "finished": true, "Ev Idx": 4273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564846262.879, "dur": 69.545,
    "args": {
      "External id": 4275,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2261, "finished": true, "Ev Idx": 4274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564846340.833, "dur": 47.729,
    "args": {
      "External id": 4276,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2262, "finished": true, "Ev Idx": 4275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564846341.340, "dur": 4.305,
    "args": {
      "External id": 4277,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2262, "finished": true, "Ev Idx": 4276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564846343.010, "dur": 1.679,
    "args": {
      "External id": 4278,"Record function id": 0, "finished": true, "Ev Idx": 4277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846343.820, "dur": 0.698,
    "args": {
      "External id": 4279,"Record function id": 0, "finished": true, "Ev Idx": 4278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564846346.124, "dur": 42.319,
    "args": {
      "External id": 4280,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2263, "finished": true, "Ev Idx": 4279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564846346.863, "dur": 2.093,
    "args": {
      "External id": 4281,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2263, "finished": true, "Ev Idx": 4280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846347.158, "dur": 1.707,
    "args": {
      "External id": 4282,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2263, "finished": true, "Ev Idx": 4281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564846349.391, "dur": 36.894,
    "args": {
      "External id": 4283,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2264, "finished": true, "Ev Idx": 4282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564846361.161, "dur": 1.079,
    "args": {
      "External id": 4284,"Record function id": 0, "finished": true, "Ev Idx": 4283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564846386.820, "dur": 1.249,
    "args": {
      "External id": 4285,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2265, "finished": true, "Ev Idx": 4284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564846390.297, "dur": 39.191,
    "args": {
      "External id": 4286,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2266, "finished": true, "Ev Idx": 4285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564846437.118, "dur": 43.601,
    "args": {
      "External id": 4287,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2267, "finished": true, "Ev Idx": 4286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564846437.557, "dur": 3.543,
    "args": {
      "External id": 4288,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2267, "finished": true, "Ev Idx": 4287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564846438.672, "dur": 1.434,
    "args": {
      "External id": 4289,"Record function id": 0, "finished": true, "Ev Idx": 4288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846439.193, "dur": 0.780,
    "args": {
      "External id": 4290,"Record function id": 0, "finished": true, "Ev Idx": 4289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564846441.474, "dur": 39.162,
    "args": {
      "External id": 4291,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2268, "finished": true, "Ev Idx": 4290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564846441.913, "dur": 2.471,
    "args": {
      "External id": 4292,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2268, "finished": true, "Ev Idx": 4291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846442.286, "dur": 2.002,
    "args": {
      "External id": 4293,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2268, "finished": true, "Ev Idx": 4292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564846444.906, "dur": 33.816,
    "args": {
      "External id": 4294,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2269, "finished": true, "Ev Idx": 4293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564846455.011, "dur": 1.260,
    "args": {
      "External id": 4295,"Record function id": 0, "finished": true, "Ev Idx": 4294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564846479.324, "dur": 0.998,
    "args": {
      "External id": 4296,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2270, "finished": true, "Ev Idx": 4295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564846483.844, "dur": 8.780,
    "args": {
      "External id": 4297,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2271, "finished": true, "Ev Idx": 4296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564846513.331, "dur": 0.511,
    "args": {
      "External id": 4298,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2272, "finished": true, "Ev Idx": 4297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564846516.054, "dur": 10.481,
    "args": {
      "External id": 4299,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2272, "finished": true, "Ev Idx": 4298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564846517.395, "dur": 0.483,
    "args": {
      "External id": 4300,"Record function id": 0, "finished": true, "Ev Idx": 4299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564846518.272, "dur": 0.100,
    "args": {
      "External id": 4301,"Record function id": 0, "finished": true, "Ev Idx": 4300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564846529.154, "dur": 13.617,
    "args": {
      "External id": 4302,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2273, "finished": true, "Ev Idx": 4301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564846546.960, "dur": 9.090,
    "args": {
      "External id": 4303,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2274, "finished": true, "Ev Idx": 4302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564846557.917, "dur": 11.711,
    "args": {
      "External id": 4304,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2275, "finished": true, "Ev Idx": 4303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564846571.499, "dur": 8.502,
    "args": {
      "External id": 4305,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2276, "finished": true, "Ev Idx": 4304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564846582.613, "dur": 0.234,
    "args": {
      "External id": 4306,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2277, "finished": true, "Ev Idx": 4305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564846583.887, "dur": 7.727,
    "args": {
      "External id": 4307,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2277, "finished": true, "Ev Idx": 4306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564846610.550, "dur": 41.078,
    "args": {
      "External id": 4308,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2278, "finished": true, "Ev Idx": 4307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564846611.107, "dur": 4.878,
    "args": {
      "External id": 4309,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2278, "finished": true, "Ev Idx": 4308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564846613.756, "dur": 1.406,
    "args": {
      "External id": 4310,"Record function id": 0, "finished": true, "Ev Idx": 4309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846614.469, "dur": 0.558,
    "args": {
      "External id": 4311,"Record function id": 0, "finished": true, "Ev Idx": 4310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564846616.335, "dur": 35.215,
    "args": {
      "External id": 4312,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2279, "finished": true, "Ev Idx": 4311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564846616.818, "dur": 1.568,
    "args": {
      "External id": 4313,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2279, "finished": true, "Ev Idx": 4312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846617.100, "dur": 1.165,
    "args": {
      "External id": 4314,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2279, "finished": true, "Ev Idx": 4313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564846618.734, "dur": 30.511,
    "args": {
      "External id": 4315,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2280, "finished": true, "Ev Idx": 4314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564846631.590, "dur": 1.237,
    "args": {
      "External id": 4316,"Record function id": 0, "finished": true, "Ev Idx": 4315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564846649.954, "dur": 1.231,
    "args": {
      "External id": 4317,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2281, "finished": true, "Ev Idx": 4316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846655.111, "dur": 1.451,
    "args": {
      "External id": 4318,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2282, "finished": true, "Ev Idx": 4317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564846658.432, "dur": 2.547,
    "args": {
      "External id": 4319,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2283, "finished": true, "Ev Idx": 4318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846659.622, "dur": 0.599,
    "args": {
      "External id": 4320,"Record function id": 0, "finished": true, "Ev Idx": 4319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564846668.435, "dur": 44.882,
    "args": {
      "External id": 4321,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2284, "finished": true, "Ev Idx": 4320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564846668.772, "dur": 2.205,
    "args": {
      "External id": 4322,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2284, "finished": true, "Ev Idx": 4321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564846669.717, "dur": 0.814,
    "args": {
      "External id": 4323,"Record function id": 0, "finished": true, "Ev Idx": 4322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846670.161, "dur": 0.271,
    "args": {
      "External id": 4324,"Record function id": 0, "finished": true, "Ev Idx": 4323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564846671.312, "dur": 41.928,
    "args": {
      "External id": 4325,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2285, "finished": true, "Ev Idx": 4324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564846671.661, "dur": 1.478,
    "args": {
      "External id": 4326,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2285, "finished": true, "Ev Idx": 4325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846671.943, "dur": 1.099,
    "args": {
      "External id": 4327,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2285, "finished": true, "Ev Idx": 4326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564846673.625, "dur": 37.610,
    "args": {
      "External id": 4328,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2286, "finished": true, "Ev Idx": 4327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564846686.849, "dur": 1.112,
    "args": {
      "External id": 4329,"Record function id": 0, "finished": true, "Ev Idx": 4328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564846711.975, "dur": 0.991,
    "args": {
      "External id": 4330,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2287, "finished": true, "Ev Idx": 4329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846715.709, "dur": 1.131,
    "args": {
      "External id": 4331,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2288, "finished": true, "Ev Idx": 4330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564846718.316, "dur": 1.935,
    "args": {
      "External id": 4332,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2289, "finished": true, "Ev Idx": 4331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846719.193, "dur": 0.456,
    "args": {
      "External id": 4333,"Record function id": 0, "finished": true, "Ev Idx": 4332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564846727.334, "dur": 37.705,
    "args": {
      "External id": 4334,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2290, "finished": true, "Ev Idx": 4333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564846727.755, "dur": 2.786,
    "args": {
      "External id": 4335,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2290, "finished": true, "Ev Idx": 4334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564846728.995, "dur": 1.144,
    "args": {
      "External id": 4336,"Record function id": 0, "finished": true, "Ev Idx": 4335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846729.611, "dur": 0.421,
    "args": {
      "External id": 4337,"Record function id": 0, "finished": true, "Ev Idx": 4336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564846730.828, "dur": 34.118,
    "args": {
      "External id": 4338,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2291, "finished": true, "Ev Idx": 4337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564846731.197, "dur": 1.315,
    "args": {
      "External id": 4339,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2291, "finished": true, "Ev Idx": 4338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846731.467, "dur": 0.955,
    "args": {
      "External id": 4340,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2291, "finished": true, "Ev Idx": 4339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564846732.811, "dur": 30.370,
    "args": {
      "External id": 4341,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2292, "finished": true, "Ev Idx": 4340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564846743.791, "dur": 1.207,
    "args": {
      "External id": 4342,"Record function id": 0, "finished": true, "Ev Idx": 4341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564846763.706, "dur": 0.958,
    "args": {
      "External id": 4343,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2293, "finished": true, "Ev Idx": 4342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564846767.394, "dur": 1.144,
    "args": {
      "External id": 4344,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2294, "finished": true, "Ev Idx": 4343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564846770.026, "dur": 1.607,
    "args": {
      "External id": 4345,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2295, "finished": true, "Ev Idx": 4344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846770.767, "dur": 0.355,
    "args": {
      "External id": 4346,"Record function id": 0, "finished": true, "Ev Idx": 4345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564846774.593, "dur": 2.293,
    "args": {
      "External id": 4347,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2296, "finished": true, "Ev Idx": 4346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846775.999, "dur": 0.342,
    "args": {
      "External id": 4348,"Record function id": 0, "finished": true, "Ev Idx": 4347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564846778.025, "dur": 1.081,
    "args": {
      "External id": 4349,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2296, "finished": true, "Ev Idx": 4348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846778.436, "dur": 0.310,
    "args": {
      "External id": 4350,"Record function id": 0, "finished": true, "Ev Idx": 4349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564846779.946, "dur": 9.182,
    "args": {
      "External id": 4351,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2296, "finished": true, "Ev Idx": 4350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564846795.125, "dur": 5.683,
    "args": {
      "External id": 4352,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2297, "finished": true, "Ev Idx": 4351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846798.488, "dur": 0.741,
    "args": {
      "External id": 4353,"Record function id": 0, "finished": true, "Ev Idx": 4352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564846803.183, "dur": 1.803,
    "args": {
      "External id": 4354,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2298, "finished": true, "Ev Idx": 4353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846804.062, "dur": 0.354,
    "args": {
      "External id": 4355,"Record function id": 0, "finished": true, "Ev Idx": 4354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564846806.775, "dur": 11.453,
    "args": {
      "External id": 4356,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2299, "finished": true, "Ev Idx": 4355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564846822.487, "dur": 16.662,
    "args": {
      "External id": 4357,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2300, "finished": true, "Ev Idx": 4356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564846842.291, "dur": 8.890,
    "args": {
      "External id": 4358,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2301, "finished": true, "Ev Idx": 4357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564846852.971, "dur": 9.322,
    "args": {
      "External id": 4359,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2302, "finished": true, "Ev Idx": 4358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564846864.159, "dur": 7.270,
    "args": {
      "External id": 4360,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2303, "finished": true, "Ev Idx": 4359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564846874.133, "dur": 2.199,
    "args": {
      "External id": 4361,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2304, "finished": true, "Ev Idx": 4360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846875.264, "dur": 0.403,
    "args": {
      "External id": 4362,"Record function id": 0, "finished": true, "Ev Idx": 4361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564846877.870, "dur": 1.867,
    "args": {
      "External id": 4363,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2305, "finished": true, "Ev Idx": 4362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846878.857, "dur": 0.331,
    "args": {
      "External id": 4364,"Record function id": 0, "finished": true, "Ev Idx": 4363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564846880.801, "dur": 8.544,
    "args": {
      "External id": 4365,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2306, "finished": true, "Ev Idx": 4364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564846891.123, "dur": 9.395,
    "args": {
      "External id": 4366,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2307, "finished": true, "Ev Idx": 4365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564846902.737, "dur": 9.582,
    "args": {
      "External id": 4367,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2308, "finished": true, "Ev Idx": 4366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564846913.581, "dur": 8.427,
    "args": {
      "External id": 4368,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2309, "finished": true, "Ev Idx": 4367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564846946.831, "dur": 2.880,
    "args": {
      "External id": 4369,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2310, "finished": true, "Ev Idx": 4368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846948.257, "dur": 0.666,
    "args": {
      "External id": 4370,"Record function id": 0, "finished": true, "Ev Idx": 4369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564846950.398, "dur": 1.686,
    "args": {
      "External id": 4371,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2311, "finished": true, "Ev Idx": 4370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846951.097, "dur": 0.453,
    "args": {
      "External id": 4372,"Record function id": 0, "finished": true, "Ev Idx": 4371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564846952.813, "dur": 2.639,
    "args": {
      "External id": 4373,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2312, "finished": true, "Ev Idx": 4372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846954.013, "dur": 0.761,
    "args": {
      "External id": 4374,"Record function id": 0, "finished": true, "Ev Idx": 4373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564846956.139, "dur": 2.975,
    "args": {
      "External id": 4375,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2313, "finished": true, "Ev Idx": 4374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846958.205, "dur": 0.381,
    "args": {
      "External id": 4376,"Record function id": 0, "finished": true, "Ev Idx": 4375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564846959.744, "dur": 2.713,
    "args": {
      "External id": 4377,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2314, "finished": true, "Ev Idx": 4376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846961.695, "dur": 0.296,
    "args": {
      "External id": 4378,"Record function id": 0, "finished": true, "Ev Idx": 4377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564846965.401, "dur": 3.106,
    "args": {
      "External id": 4379,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2315, "finished": true, "Ev Idx": 4378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846967.336, "dur": 0.392,
    "args": {
      "External id": 4380,"Record function id": 0, "finished": true, "Ev Idx": 4379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564846970.833, "dur": 20.487,
    "args": {
      "External id": 4381,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2316, "finished": true, "Ev Idx": 4380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564846973.422, "dur": 15.751,
    "args": {
      "External id": 4382,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2316, "finished": true, "Ev Idx": 4381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564846975.012, "dur": 2.901,
    "args": {
      "External id": 4383,"Record function id": 0, "finished": true, "Ev Idx": 4382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564846975.989, "dur": 1.661,
    "args": {
      "External id": 4384,"Record function id": 0, "finished": true, "Ev Idx": 4383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564846978.627, "dur": 9.946,
    "args": {
      "External id": 4385,"Record function id": 0, "finished": true, "Ev Idx": 4384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564846989.769, "dur": 1.260,
    "args": {
      "External id": 4386,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2317, "finished": true, "Ev Idx": 4385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564846994.514, "dur": 2.815,
    "args": {
      "External id": 4387,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2318, "finished": true, "Ev Idx": 4386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564846996.296, "dur": 0.446,
    "args": {
      "External id": 4388,"Record function id": 0, "finished": true, "Ev Idx": 4387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564846998.105, "dur": 2.968,
    "args": {
      "External id": 4389,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2319, "finished": true, "Ev Idx": 4388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847000.171, "dur": 0.443,
    "args": {
      "External id": 4390,"Record function id": 0, "finished": true, "Ev Idx": 4389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564847001.618, "dur": 2.259,
    "args": {
      "External id": 4391,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2320, "finished": true, "Ev Idx": 4390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847003.120, "dur": 0.334,
    "args": {
      "External id": 4392,"Record function id": 0, "finished": true, "Ev Idx": 4391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564847004.454, "dur": 1.466,
    "args": {
      "External id": 4393,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2321, "finished": true, "Ev Idx": 4392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847005.107, "dur": 0.336,
    "args": {
      "External id": 4394,"Record function id": 0, "finished": true, "Ev Idx": 4393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564847006.520, "dur": 1.501,
    "args": {
      "External id": 4395,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2322, "finished": true, "Ev Idx": 4394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847007.153, "dur": 0.416,
    "args": {
      "External id": 4396,"Record function id": 0, "finished": true, "Ev Idx": 4395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564847009.616, "dur": 1.634,
    "args": {
      "External id": 4397,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2323, "finished": true, "Ev Idx": 4396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847010.374, "dur": 0.444,
    "args": {
      "External id": 4398,"Record function id": 0, "finished": true, "Ev Idx": 4397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847012.739, "dur": 12.302,
    "args": {
      "External id": 4399,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2324, "finished": true, "Ev Idx": 4398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564847013.635, "dur": 9.397,
    "args": {
      "External id": 4400,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2324, "finished": true, "Ev Idx": 4399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564847014.443, "dur": 1.645,
    "args": {
      "External id": 4401,"Record function id": 0, "finished": true, "Ev Idx": 4400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847014.774, "dur": 1.179,
    "args": {
      "External id": 4402,"Record function id": 0, "finished": true, "Ev Idx": 4401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564847016.414, "dur": 6.276,
    "args": {
      "External id": 4403,"Record function id": 0, "finished": true, "Ev Idx": 4402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564847023.411, "dur": 1.455,
    "args": {
      "External id": 4404,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2325, "finished": true, "Ev Idx": 4403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847027.229, "dur": 1.838,
    "args": {
      "External id": 4405,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2326, "finished": true, "Ev Idx": 4404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847028.182, "dur": 0.432,
    "args": {
      "External id": 4406,"Record function id": 0, "finished": true, "Ev Idx": 4405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564847030.576, "dur": 47.662,
    "args": {
      "External id": 4407,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2327, "finished": true, "Ev Idx": 4406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564847031.298, "dur": 1.722,
    "args": {
      "External id": 4408,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2327, "finished": true, "Ev Idx": 4407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847032.243, "dur": 0.336,
    "args": {
      "External id": 4409,"Record function id": 0, "finished": true, "Ev Idx": 4408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847033.326, "dur": 3.816,
    "args": {
      "External id": 4410,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2328, "finished": true, "Ev Idx": 4409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564847035.084, "dur": 1.898,
    "args": {
      "External id": 4411,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2328, "finished": true, "Ev Idx": 4410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564847037.608, "dur": 1.334,
    "args": {
      "External id": 4412,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2329, "finished": true, "Ev Idx": 4411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847038.320, "dur": 0.278,
    "args": {
      "External id": 4413,"Record function id": 0, "finished": true, "Ev Idx": 4412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847039.558, "dur": 2.461,
    "args": {
      "External id": 4414,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2330, "finished": true, "Ev Idx": 4413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564847040.395, "dur": 1.488,
    "args": {
      "External id": 4415,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2330, "finished": true, "Ev Idx": 4414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564847042.759, "dur": 33.279,
    "args": {
      "External id": 4416,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2331, "finished": true, "Ev Idx": 4415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847057.919, "dur": 0.950,
    "args": {
      "External id": 4417,"Record function id": 0, "finished": true, "Ev Idx": 4416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564847076.686, "dur": 1.096,
    "args": {
      "External id": 4418,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2332, "finished": true, "Ev Idx": 4417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564847081.389, "dur": 9.486,
    "args": {
      "External id": 4419,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2333, "finished": true, "Ev Idx": 4418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564847093.938, "dur": 1.812,
    "args": {
      "External id": 4420,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2334, "finished": true, "Ev Idx": 4419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847094.699, "dur": 0.404,
    "args": {
      "External id": 4421,"Record function id": 0, "finished": true, "Ev Idx": 4420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564847096.387, "dur": 1.322,
    "args": {
      "External id": 4422,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2334, "finished": true, "Ev Idx": 4421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847097.032, "dur": 0.308,
    "args": {
      "External id": 4423,"Record function id": 0, "finished": true, "Ev Idx": 4422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564847098.348, "dur": 1.076,
    "args": {
      "External id": 4424,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2334, "finished": true, "Ev Idx": 4423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847098.821, "dur": 0.271,
    "args": {
      "External id": 4425,"Record function id": 0, "finished": true, "Ev Idx": 4424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564847101.072, "dur": 8.569,
    "args": {
      "External id": 4426,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2334, "finished": true, "Ev Idx": 4425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564847115.221, "dur": 14.039,
    "args": {
      "External id": 4427,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2335, "finished": true, "Ev Idx": 4426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564847116.183, "dur": 0.386,
    "args": {
      "External id": 4428,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2335, "finished": true, "Ev Idx": 4427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564847117.267, "dur": 11.692,
    "args": {
      "External id": 4429,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2335, "finished": true, "Ev Idx": 4428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564847132.349, "dur": 0.192,
    "args": {
      "External id": 4430,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2336, "finished": true, "Ev Idx": 4429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564847143.306, "dur": 0.666,
    "args": {
      "External id": 4431,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2336, "finished": true, "Ev Idx": 4430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564847146.029, "dur": 40.535,
    "args": {
      "External id": 4432,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2336, "finished": true, "Ev Idx": 4431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564847146.513, "dur": 2.063,
    "args": {
      "External id": 4433,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2336, "finished": true, "Ev Idx": 4432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847147.540, "dur": 0.397,
    "args": {
      "External id": 4434,"Record function id": 0, "finished": true, "Ev Idx": 4433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847148.983, "dur": 3.428,
    "args": {
      "External id": 4435,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2337, "finished": true, "Ev Idx": 4434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847149.399, "dur": 2.880,
    "args": {
      "External id": 4436,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2337, "finished": true, "Ev Idx": 4435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564847152.819, "dur": 1.363,
    "args": {
      "External id": 4437,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2338, "finished": true, "Ev Idx": 4436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847153.589, "dur": 0.238,
    "args": {
      "External id": 4438,"Record function id": 0, "finished": true, "Ev Idx": 4437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847154.456, "dur": 1.382,
    "args": {
      "External id": 4439,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2339, "finished": true, "Ev Idx": 4438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847154.721, "dur": 1.045,
    "args": {
      "External id": 4440,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2339, "finished": true, "Ev Idx": 4439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564847156.385, "dur": 28.029,
    "args": {
      "External id": 4441,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2340, "finished": true, "Ev Idx": 4440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847167.209, "dur": 1.146,
    "args": {
      "External id": 4442,"Record function id": 0, "finished": true, "Ev Idx": 4441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564847185.176, "dur": 1.077,
    "args": {
      "External id": 4443,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2341, "finished": true, "Ev Idx": 4442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847188.710, "dur": 2.098,
    "args": {
      "External id": 4444,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2342, "finished": true, "Ev Idx": 4443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847189.889, "dur": 0.398,
    "args": {
      "External id": 4445,"Record function id": 0, "finished": true, "Ev Idx": 4444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564847193.400, "dur": 11.959,
    "args": {
      "External id": 4446,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2343, "finished": true, "Ev Idx": 4445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564847194.033, "dur": 11.166,
    "args": {
      "External id": 4447,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2343, "finished": true, "Ev Idx": 4446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564847195.026, "dur": 2.148,
    "args": {
      "External id": 4448,"Record function id": 0, "finished": true, "Ev Idx": 4447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847195.679, "dur": 1.320,
    "args": {
      "External id": 4449,"Record function id": 0, "finished": true, "Ev Idx": 4448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564847197.643, "dur": 7.147,
    "args": {
      "External id": 4450,"Record function id": 0, "finished": true, "Ev Idx": 4449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847210.801, "dur": 1.686,
    "args": {
      "External id": 4451,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2344, "finished": true, "Ev Idx": 4450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847211.140, "dur": 1.241,
    "args": {
      "External id": 4452,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2344, "finished": true, "Ev Idx": 4451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564847225.650, "dur": 36.876,
    "args": {
      "External id": 4453,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2345, "finished": true, "Ev Idx": 4452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564847226.064, "dur": 2.759,
    "args": {
      "External id": 4454,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2345, "finished": true, "Ev Idx": 4453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847227.309, "dur": 0.954,
    "args": {
      "External id": 4455,"Record function id": 0, "finished": true, "Ev Idx": 4454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847227.750, "dur": 0.401,
    "args": {
      "External id": 4456,"Record function id": 0, "finished": true, "Ev Idx": 4455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564847229.156, "dur": 33.233,
    "args": {
      "External id": 4457,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2346, "finished": true, "Ev Idx": 4456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847229.571, "dur": 1.697,
    "args": {
      "External id": 4458,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2346, "finished": true, "Ev Idx": 4457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847229.986, "dur": 1.198,
    "args": {
      "External id": 4459,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2346, "finished": true, "Ev Idx": 4458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564847231.867, "dur": 28.426,
    "args": {
      "External id": 4460,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2347, "finished": true, "Ev Idx": 4459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847245.003, "dur": 1.054,
    "args": {
      "External id": 4461,"Record function id": 0, "finished": true, "Ev Idx": 4460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564847260.877, "dur": 1.188,
    "args": {
      "External id": 4462,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2348, "finished": true, "Ev Idx": 4461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564847267.132, "dur": 9.176,
    "args": {
      "External id": 4463,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2349, "finished": true, "Ev Idx": 4462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564847285.398, "dur": 0.289,
    "args": {
      "External id": 4464,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2350, "finished": true, "Ev Idx": 4463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564847287.652, "dur": 11.968,
    "args": {
      "External id": 4465,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2350, "finished": true, "Ev Idx": 4464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564847289.307, "dur": 0.409,
    "args": {
      "External id": 4466,"Record function id": 0, "finished": true, "Ev Idx": 4465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564847290.162, "dur": 0.375,
    "args": {
      "External id": 4467,"Record function id": 0, "finished": true, "Ev Idx": 4466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564847302.493, "dur": 12.283,
    "args": {
      "External id": 4468,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2351, "finished": true, "Ev Idx": 4467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564847318.427, "dur": 8.771,
    "args": {
      "External id": 4469,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2352, "finished": true, "Ev Idx": 4468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564847329.223, "dur": 8.386,
    "args": {
      "External id": 4470,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2353, "finished": true, "Ev Idx": 4469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564847339.255, "dur": 8.191,
    "args": {
      "External id": 4471,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2354, "finished": true, "Ev Idx": 4470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564847349.854, "dur": 0.266,
    "args": {
      "External id": 4472,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2355, "finished": true, "Ev Idx": 4471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564847351.014, "dur": 68.487,
    "args": {
      "External id": 4473,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2355, "finished": true, "Ev Idx": 4472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564847435.610, "dur": 56.668,
    "args": {
      "External id": 4474,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2356, "finished": true, "Ev Idx": 4473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564847436.408, "dur": 5.138,
    "args": {
      "External id": 4475,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2356, "finished": true, "Ev Idx": 4474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847439.187, "dur": 1.470,
    "args": {
      "External id": 4476,"Record function id": 0, "finished": true, "Ev Idx": 4475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847439.953, "dur": 0.528,
    "args": {
      "External id": 4477,"Record function id": 0, "finished": true, "Ev Idx": 4476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564847442.025, "dur": 50.126,
    "args": {
      "External id": 4478,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2357, "finished": true, "Ev Idx": 4477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847442.766, "dur": 2.129,
    "args": {
      "External id": 4479,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2357, "finished": true, "Ev Idx": 4478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847443.093, "dur": 1.667,
    "args": {
      "External id": 4480,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2357, "finished": true, "Ev Idx": 4479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564847445.266, "dur": 44.749,
    "args": {
      "External id": 4481,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2358, "finished": true, "Ev Idx": 4480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847461.003, "dur": 1.270,
    "args": {
      "External id": 4482,"Record function id": 0, "finished": true, "Ev Idx": 4481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564847490.708, "dur": 1.089,
    "args": {
      "External id": 4483,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2359, "finished": true, "Ev Idx": 4482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564847500.669, "dur": 10.636,
    "args": {
      "External id": 4484,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2360, "finished": true, "Ev Idx": 4483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564847518.797, "dur": 80.565,
    "args": {
      "External id": 4485,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2361, "finished": true, "Ev Idx": 4484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564847519.107, "dur": 2.645,
    "args": {
      "External id": 4486,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2361, "finished": true, "Ev Idx": 4485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847520.069, "dur": 1.095,
    "args": {
      "External id": 4487,"Record function id": 0, "finished": true, "Ev Idx": 4486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847520.575, "dur": 0.432,
    "args": {
      "External id": 4488,"Record function id": 0, "finished": true, "Ev Idx": 4487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564847522.077, "dur": 77.174,
    "args": {
      "External id": 4489,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2362, "finished": true, "Ev Idx": 4488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847522.645, "dur": 2.173,
    "args": {
      "External id": 4490,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2362, "finished": true, "Ev Idx": 4489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847523.220, "dur": 1.488,
    "args": {
      "External id": 4491,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2362, "finished": true, "Ev Idx": 4490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564847525.193, "dur": 71.748,
    "args": {
      "External id": 4492,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2363, "finished": true, "Ev Idx": 4491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847571.743, "dur": 1.136,
    "args": {
      "External id": 4493,"Record function id": 0, "finished": true, "Ev Idx": 4492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564847597.676, "dur": 1.219,
    "args": {
      "External id": 4494,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2364, "finished": true, "Ev Idx": 4493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564847601.719, "dur": 8.629,
    "args": {
      "External id": 4495,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2365, "finished": true, "Ev Idx": 4494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564847617.203, "dur": 41.901,
    "args": {
      "External id": 4496,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2366, "finished": true, "Ev Idx": 4495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564847617.835, "dur": 2.671,
    "args": {
      "External id": 4497,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2366, "finished": true, "Ev Idx": 4496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847618.865, "dur": 1.067,
    "args": {
      "External id": 4498,"Record function id": 0, "finished": true, "Ev Idx": 4497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847619.336, "dur": 0.435,
    "args": {
      "External id": 4499,"Record function id": 0, "finished": true, "Ev Idx": 4498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564847620.845, "dur": 38.161,
    "args": {
      "External id": 4500,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2367, "finished": true, "Ev Idx": 4499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847621.226, "dur": 1.912,
    "args": {
      "External id": 4501,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2367, "finished": true, "Ev Idx": 4500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847621.665, "dur": 1.379,
    "args": {
      "External id": 4502,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2367, "finished": true, "Ev Idx": 4501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564847623.524, "dur": 33.735,
    "args": {
      "External id": 4503,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2368, "finished": true, "Ev Idx": 4502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847635.219, "dur": 1.269,
    "args": {
      "External id": 4504,"Record function id": 0, "finished": true, "Ev Idx": 4503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564847657.815, "dur": 0.912,
    "args": {
      "External id": 4505,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2369, "finished": true, "Ev Idx": 4504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564847662.106, "dur": 8.678,
    "args": {
      "External id": 4506,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2370, "finished": true, "Ev Idx": 4505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564847691.035, "dur": 0.225,
    "args": {
      "External id": 4507,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2371, "finished": true, "Ev Idx": 4506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564847693.083, "dur": 10.091,
    "args": {
      "External id": 4508,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2371, "finished": true, "Ev Idx": 4507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564847694.517, "dur": 0.549,
    "args": {
      "External id": 4509,"Record function id": 0, "finished": true, "Ev Idx": 4508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564847695.432, "dur": 0.066,
    "args": {
      "External id": 4510,"Record function id": 0, "finished": true, "Ev Idx": 4509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564847705.709, "dur": 13.237,
    "args": {
      "External id": 4511,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2372, "finished": true, "Ev Idx": 4510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564847722.661, "dur": 7.600,
    "args": {
      "External id": 4512,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2373, "finished": true, "Ev Idx": 4511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564847731.900, "dur": 10.096,
    "args": {
      "External id": 4513,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2374, "finished": true, "Ev Idx": 4512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564847743.377, "dur": 8.973,
    "args": {
      "External id": 4514,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2375, "finished": true, "Ev Idx": 4513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564847754.683, "dur": 0.103,
    "args": {
      "External id": 4515,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2376, "finished": true, "Ev Idx": 4514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564847755.718, "dur": 8.933,
    "args": {
      "External id": 4516,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2376, "finished": true, "Ev Idx": 4515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564847781.956, "dur": 76.616,
    "args": {
      "External id": 4517,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2377, "finished": true, "Ev Idx": 4516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564847782.491, "dur": 3.794,
    "args": {
      "External id": 4518,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2377, "finished": true, "Ev Idx": 4517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847783.950, "dur": 1.482,
    "args": {
      "External id": 4519,"Record function id": 0, "finished": true, "Ev Idx": 4518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847784.797, "dur": 0.512,
    "args": {
      "External id": 4520,"Record function id": 0, "finished": true, "Ev Idx": 4519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564847786.789, "dur": 71.701,
    "args": {
      "External id": 4521,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2378, "finished": true, "Ev Idx": 4520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847787.414, "dur": 2.563,
    "args": {
      "External id": 4522,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2378, "finished": true, "Ev Idx": 4521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847787.820, "dur": 2.025,
    "args": {
      "External id": 4523,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2378, "finished": true, "Ev Idx": 4522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564847790.457, "dur": 65.550,
    "args": {
      "External id": 4524,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2379, "finished": true, "Ev Idx": 4523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847838.471, "dur": 1.186,
    "args": {
      "External id": 4525,"Record function id": 0, "finished": true, "Ev Idx": 4524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564847856.814, "dur": 1.314,
    "args": {
      "External id": 4526,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2380, "finished": true, "Ev Idx": 4525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847862.150, "dur": 1.461,
    "args": {
      "External id": 4527,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2381, "finished": true, "Ev Idx": 4526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847865.522, "dur": 2.545,
    "args": {
      "External id": 4528,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2382, "finished": true, "Ev Idx": 4527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847866.650, "dur": 0.569,
    "args": {
      "External id": 4529,"Record function id": 0, "finished": true, "Ev Idx": 4528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564847875.390, "dur": 41.519,
    "args": {
      "External id": 4530,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2383, "finished": true, "Ev Idx": 4529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564847875.919, "dur": 2.569,
    "args": {
      "External id": 4531,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2383, "finished": true, "Ev Idx": 4530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847876.806, "dur": 1.201,
    "args": {
      "External id": 4532,"Record function id": 0, "finished": true, "Ev Idx": 4531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847877.628, "dur": 0.258,
    "args": {
      "External id": 4533,"Record function id": 0, "finished": true, "Ev Idx": 4532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564847878.805, "dur": 38.023,
    "args": {
      "External id": 4534,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2384, "finished": true, "Ev Idx": 4533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847879.254, "dur": 1.558,
    "args": {
      "External id": 4535,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2384, "finished": true, "Ev Idx": 4534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847879.532, "dur": 1.172,
    "args": {
      "External id": 4536,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2384, "finished": true, "Ev Idx": 4535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564847881.181, "dur": 33.719,
    "args": {
      "External id": 4537,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2385, "finished": true, "Ev Idx": 4536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847891.791, "dur": 1.165,
    "args": {
      "External id": 4538,"Record function id": 0, "finished": true, "Ev Idx": 4537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564847915.451, "dur": 1.102,
    "args": {
      "External id": 4539,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2386, "finished": true, "Ev Idx": 4538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847919.181, "dur": 1.125,
    "args": {
      "External id": 4540,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2387, "finished": true, "Ev Idx": 4539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847921.685, "dur": 1.817,
    "args": {
      "External id": 4541,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2388, "finished": true, "Ev Idx": 4540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847922.472, "dur": 0.408,
    "args": {
      "External id": 4542,"Record function id": 0, "finished": true, "Ev Idx": 4541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564847930.547, "dur": 39.999,
    "args": {
      "External id": 4543,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2389, "finished": true, "Ev Idx": 4542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564847930.874, "dur": 2.023,
    "args": {
      "External id": 4544,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2389, "finished": true, "Ev Idx": 4543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847931.636, "dur": 0.815,
    "args": {
      "External id": 4545,"Record function id": 0, "finished": true, "Ev Idx": 4544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847932.058, "dur": 0.273,
    "args": {
      "External id": 4546,"Record function id": 0, "finished": true, "Ev Idx": 4545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564847933.180, "dur": 37.245,
    "args": {
      "External id": 4547,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2390, "finished": true, "Ev Idx": 4546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564847933.682, "dur": 1.304,
    "args": {
      "External id": 4548,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2390, "finished": true, "Ev Idx": 4547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847934.101, "dur": 0.788,
    "args": {
      "External id": 4549,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2390, "finished": true, "Ev Idx": 4548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564847935.492, "dur": 32.882,
    "args": {
      "External id": 4550,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2391, "finished": true, "Ev Idx": 4549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564847946.725, "dur": 1.151,
    "args": {
      "External id": 4551,"Record function id": 0, "finished": true, "Ev Idx": 4550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564847969.074, "dur": 1.062,
    "args": {
      "External id": 4552,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2392, "finished": true, "Ev Idx": 4551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564847972.853, "dur": 1.159,
    "args": {
      "External id": 4553,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2393, "finished": true, "Ev Idx": 4552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564847975.498, "dur": 1.773,
    "args": {
      "External id": 4554,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2394, "finished": true, "Ev Idx": 4553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847976.213, "dur": 0.513,
    "args": {
      "External id": 4555,"Record function id": 0, "finished": true, "Ev Idx": 4554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564847979.943, "dur": 2.273,
    "args": {
      "External id": 4556,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2395, "finished": true, "Ev Idx": 4555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847981.401, "dur": 0.325,
    "args": {
      "External id": 4557,"Record function id": 0, "finished": true, "Ev Idx": 4556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564847983.287, "dur": 1.199,
    "args": {
      "External id": 4558,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2395, "finished": true, "Ev Idx": 4557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564847983.880, "dur": 0.298,
    "args": {
      "External id": 4559,"Record function id": 0, "finished": true, "Ev Idx": 4558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564847985.355, "dur": 11.515,
    "args": {
      "External id": 4560,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2395, "finished": true, "Ev Idx": 4559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848002.038, "dur": 5.479,
    "args": {
      "External id": 4561,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2396, "finished": true, "Ev Idx": 4560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848005.338, "dur": 0.776,
    "args": {
      "External id": 4562,"Record function id": 0, "finished": true, "Ev Idx": 4561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848009.712, "dur": 2.328,
    "args": {
      "External id": 4563,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2397, "finished": true, "Ev Idx": 4562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848010.710, "dur": 0.567,
    "args": {
      "External id": 4564,"Record function id": 0, "finished": true, "Ev Idx": 4563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564848013.945, "dur": 10.308,
    "args": {
      "External id": 4565,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2398, "finished": true, "Ev Idx": 4564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564848028.620, "dur": 16.278,
    "args": {
      "External id": 4566,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2399, "finished": true, "Ev Idx": 4565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564848047.807, "dur": 9.879,
    "args": {
      "External id": 4567,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2400, "finished": true, "Ev Idx": 4566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564848059.247, "dur": 7.981,
    "args": {
      "External id": 4568,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2401, "finished": true, "Ev Idx": 4567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564848069.446, "dur": 8.343,
    "args": {
      "External id": 4569,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2402, "finished": true, "Ev Idx": 4568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848080.352, "dur": 3.124,
    "args": {
      "External id": 4570,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2403, "finished": true, "Ev Idx": 4569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848082.322, "dur": 0.456,
    "args": {
      "External id": 4571,"Record function id": 0, "finished": true, "Ev Idx": 4570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848085.039, "dur": 2.244,
    "args": {
      "External id": 4572,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2404, "finished": true, "Ev Idx": 4571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848086.420, "dur": 0.339,
    "args": {
      "External id": 4573,"Record function id": 0, "finished": true, "Ev Idx": 4572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564848088.301, "dur": 7.186,
    "args": {
      "External id": 4574,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2405, "finished": true, "Ev Idx": 4573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564848097.273, "dur": 11.269,
    "args": {
      "External id": 4575,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2406, "finished": true, "Ev Idx": 4574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564848110.581, "dur": 7.503,
    "args": {
      "External id": 4576,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2407, "finished": true, "Ev Idx": 4575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564848119.995, "dur": 6.297,
    "args": {
      "External id": 4577,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2408, "finished": true, "Ev Idx": 4576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848149.759, "dur": 3.346,
    "args": {
      "External id": 4578,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2409, "finished": true, "Ev Idx": 4577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848151.887, "dur": 0.501,
    "args": {
      "External id": 4579,"Record function id": 0, "finished": true, "Ev Idx": 4578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848153.717, "dur": 2.583,
    "args": {
      "External id": 4580,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2410, "finished": true, "Ev Idx": 4579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848155.356, "dur": 0.434,
    "args": {
      "External id": 4581,"Record function id": 0, "finished": true, "Ev Idx": 4580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564848157.132, "dur": 2.332,
    "args": {
      "External id": 4582,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2411, "finished": true, "Ev Idx": 4581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848158.218, "dur": 0.632,
    "args": {
      "External id": 4583,"Record function id": 0, "finished": true, "Ev Idx": 4582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848160.161, "dur": 1.723,
    "args": {
      "External id": 4584,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2412, "finished": true, "Ev Idx": 4583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848160.980, "dur": 0.374,
    "args": {
      "External id": 4585,"Record function id": 0, "finished": true, "Ev Idx": 4584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848162.564, "dur": 1.669,
    "args": {
      "External id": 4586,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2413, "finished": true, "Ev Idx": 4585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848163.418, "dur": 0.328,
    "args": {
      "External id": 4587,"Record function id": 0, "finished": true, "Ev Idx": 4586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564848166.970, "dur": 3.265,
    "args": {
      "External id": 4588,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2414, "finished": true, "Ev Idx": 4587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848169.125, "dur": 0.406,
    "args": {
      "External id": 4589,"Record function id": 0, "finished": true, "Ev Idx": 4588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848172.414, "dur": 20.228,
    "args": {
      "External id": 4590,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2415, "finished": true, "Ev Idx": 4589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564848175.128, "dur": 15.558,
    "args": {
      "External id": 4591,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2415, "finished": true, "Ev Idx": 4590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564848177.281, "dur": 2.798,
    "args": {
      "External id": 4592,"Record function id": 0, "finished": true, "Ev Idx": 4591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564848178.239, "dur": 1.624,
    "args": {
      "External id": 4593,"Record function id": 0, "finished": true, "Ev Idx": 4592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564848180.727, "dur": 9.304,
    "args": {
      "External id": 4594,"Record function id": 0, "finished": true, "Ev Idx": 4593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564848191.200, "dur": 1.212,
    "args": {
      "External id": 4595,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2416, "finished": true, "Ev Idx": 4594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848195.540, "dur": 2.004,
    "args": {
      "External id": 4596,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2417, "finished": true, "Ev Idx": 4595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848196.533, "dur": 0.437,
    "args": {
      "External id": 4597,"Record function id": 0, "finished": true, "Ev Idx": 4596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848198.113, "dur": 1.487,
    "args": {
      "External id": 4598,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2418, "finished": true, "Ev Idx": 4597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848198.861, "dur": 0.272,
    "args": {
      "External id": 4599,"Record function id": 0, "finished": true, "Ev Idx": 4598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564848200.144, "dur": 1.433,
    "args": {
      "External id": 4600,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2419, "finished": true, "Ev Idx": 4599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848200.840, "dur": 0.317,
    "args": {
      "External id": 4601,"Record function id": 0, "finished": true, "Ev Idx": 4600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848202.363, "dur": 1.662,
    "args": {
      "External id": 4602,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2420, "finished": true, "Ev Idx": 4601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848203.241, "dur": 0.321,
    "args": {
      "External id": 4603,"Record function id": 0, "finished": true, "Ev Idx": 4602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848204.753, "dur": 2.269,
    "args": {
      "External id": 4604,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2421, "finished": true, "Ev Idx": 4603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848206.265, "dur": 0.301,
    "args": {
      "External id": 4605,"Record function id": 0, "finished": true, "Ev Idx": 4604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564848208.694, "dur": 2.366,
    "args": {
      "External id": 4606,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2422, "finished": true, "Ev Idx": 4605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848210.287, "dur": 0.358,
    "args": {
      "External id": 4607,"Record function id": 0, "finished": true, "Ev Idx": 4606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848212.530, "dur": 44.404,
    "args": {
      "External id": 4608,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2423, "finished": true, "Ev Idx": 4607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564848245.788, "dur": 9.617,
    "args": {
      "External id": 4609,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2423, "finished": true, "Ev Idx": 4608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564848246.503, "dur": 1.682,
    "args": {
      "External id": 4610,"Record function id": 0, "finished": true, "Ev Idx": 4609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564848246.870, "dur": 1.191,
    "args": {
      "External id": 4611,"Record function id": 0, "finished": true, "Ev Idx": 4610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564848248.526, "dur": 6.521,
    "args": {
      "External id": 4612,"Record function id": 0, "finished": true, "Ev Idx": 4611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564848255.785, "dur": 0.964,
    "args": {
      "External id": 4613,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2424, "finished": true, "Ev Idx": 4612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564848258.936, "dur": 1.663,
    "args": {
      "External id": 4614,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2425, "finished": true, "Ev Idx": 4613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848259.754, "dur": 0.396,
    "args": {
      "External id": 4615,"Record function id": 0, "finished": true, "Ev Idx": 4614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564848262.240, "dur": 45.329,
    "args": {
      "External id": 4616,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2426, "finished": true, "Ev Idx": 4615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564848262.891, "dur": 3.098,
    "args": {
      "External id": 4617,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2426, "finished": true, "Ev Idx": 4616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848265.093, "dur": 0.433,
    "args": {
      "External id": 4618,"Record function id": 0, "finished": true, "Ev Idx": 4617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848266.299, "dur": 4.963,
    "args": {
      "External id": 4619,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2427, "finished": true, "Ev Idx": 4618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564848268.215, "dur": 2.934,
    "args": {
      "External id": 4620,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2427, "finished": true, "Ev Idx": 4619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564848271.735, "dur": 1.283,
    "args": {
      "External id": 4621,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2428, "finished": true, "Ev Idx": 4620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848272.379, "dur": 0.284,
    "args": {
      "External id": 4622,"Record function id": 0, "finished": true, "Ev Idx": 4621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848273.295, "dur": 1.817,
    "args": {
      "External id": 4623,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2429, "finished": true, "Ev Idx": 4622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564848274.173, "dur": 0.828,
    "args": {
      "External id": 4624,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2429, "finished": true, "Ev Idx": 4623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564848275.777, "dur": 29.758,
    "args": {
      "External id": 4625,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2430, "finished": true, "Ev Idx": 4624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564848287.982, "dur": 0.982,
    "args": {
      "External id": 4626,"Record function id": 0, "finished": true, "Ev Idx": 4625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564848306.152, "dur": 1.103,
    "args": {
      "External id": 4627,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2431, "finished": true, "Ev Idx": 4626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564848310.774, "dur": 9.037,
    "args": {
      "External id": 4628,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2432, "finished": true, "Ev Idx": 4627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848322.946, "dur": 1.703,
    "args": {
      "External id": 4629,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2433, "finished": true, "Ev Idx": 4628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848323.685, "dur": 0.379,
    "args": {
      "External id": 4630,"Record function id": 0, "finished": true, "Ev Idx": 4629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848325.279, "dur": 1.166,
    "args": {
      "External id": 4631,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2433, "finished": true, "Ev Idx": 4630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848325.802, "dur": 0.303,
    "args": {
      "External id": 4632,"Record function id": 0, "finished": true, "Ev Idx": 4631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564848327.107, "dur": 0.950,
    "args": {
      "External id": 4633,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2433, "finished": true, "Ev Idx": 4632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848327.528, "dur": 0.219,
    "args": {
      "External id": 4634,"Record function id": 0, "finished": true, "Ev Idx": 4633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564848329.801, "dur": 7.195,
    "args": {
      "External id": 4635,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2433, "finished": true, "Ev Idx": 4634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564848342.079, "dur": 12.938,
    "args": {
      "External id": 4636,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2434, "finished": true, "Ev Idx": 4635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564848343.013, "dur": 0.377,
    "args": {
      "External id": 4637,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2434, "finished": true, "Ev Idx": 4636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564848344.029, "dur": 10.707,
    "args": {
      "External id": 4638,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2434, "finished": true, "Ev Idx": 4637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564848358.107, "dur": 0.215,
    "args": {
      "External id": 4639,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2435, "finished": true, "Ev Idx": 4638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564848368.945, "dur": 0.503,
    "args": {
      "External id": 4640,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2435, "finished": true, "Ev Idx": 4639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564848371.288, "dur": 37.271,
    "args": {
      "External id": 4641,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2435, "finished": true, "Ev Idx": 4640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564848371.824, "dur": 1.840,
    "args": {
      "External id": 4642,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2435, "finished": true, "Ev Idx": 4641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848372.824, "dur": 0.377,
    "args": {
      "External id": 4643,"Record function id": 0, "finished": true, "Ev Idx": 4642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848374.015, "dur": 2.508,
    "args": {
      "External id": 4644,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2436, "finished": true, "Ev Idx": 4643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564848374.341, "dur": 2.067,
    "args": {
      "External id": 4645,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2436, "finished": true, "Ev Idx": 4644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564848376.905, "dur": 2.026,
    "args": {
      "External id": 4646,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2437, "finished": true, "Ev Idx": 4645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848378.293, "dur": 0.265,
    "args": {
      "External id": 4647,"Record function id": 0, "finished": true, "Ev Idx": 4646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848379.242, "dur": 2.071,
    "args": {
      "External id": 4648,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2438, "finished": true, "Ev Idx": 4647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564848379.490, "dur": 1.764,
    "args": {
      "External id": 4649,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2438, "finished": true, "Ev Idx": 4648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564848381.682, "dur": 25.002,
    "args": {
      "External id": 4650,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2439, "finished": true, "Ev Idx": 4649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564848390.597, "dur": 1.021,
    "args": {
      "External id": 4651,"Record function id": 0, "finished": true, "Ev Idx": 4650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564848407.259, "dur": 0.998,
    "args": {
      "External id": 4652,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2440, "finished": true, "Ev Idx": 4651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564848410.548, "dur": 1.891,
    "args": {
      "External id": 4653,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2441, "finished": true, "Ev Idx": 4652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848411.517, "dur": 0.403,
    "args": {
      "External id": 4654,"Record function id": 0, "finished": true, "Ev Idx": 4653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564848415.012, "dur": 12.265,
    "args": {
      "External id": 4655,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2442, "finished": true, "Ev Idx": 4654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564848415.612, "dur": 11.527,
    "args": {
      "External id": 4656,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2442, "finished": true, "Ev Idx": 4655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564848416.562, "dur": 1.953,
    "args": {
      "External id": 4657,"Record function id": 0, "finished": true, "Ev Idx": 4656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564848416.907, "dur": 1.464,
    "args": {
      "External id": 4658,"Record function id": 0, "finished": true, "Ev Idx": 4657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564848418.885, "dur": 7.808,
    "args": {
      "External id": 4659,"Record function id": 0, "finished": true, "Ev Idx": 4658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848432.783, "dur": 1.636,
    "args": {
      "External id": 4660,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2443, "finished": true, "Ev Idx": 4659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564848433.076, "dur": 1.258,
    "args": {
      "External id": 4661,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2443, "finished": true, "Ev Idx": 4660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564848446.913, "dur": 34.787,
    "args": {
      "External id": 4662,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2444, "finished": true, "Ev Idx": 4661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564848447.494, "dur": 2.784,
    "args": {
      "External id": 4663,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2444, "finished": true, "Ev Idx": 4662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564848448.807, "dur": 0.936,
    "args": {
      "External id": 4664,"Record function id": 0, "finished": true, "Ev Idx": 4663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848449.274, "dur": 0.368,
    "args": {
      "External id": 4665,"Record function id": 0, "finished": true, "Ev Idx": 4664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564848450.625, "dur": 30.935,
    "args": {
      "External id": 4666,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2445, "finished": true, "Ev Idx": 4665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848451.003, "dur": 2.108,
    "args": {
      "External id": 4667,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2445, "finished": true, "Ev Idx": 4666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564848451.252, "dur": 1.770,
    "args": {
      "External id": 4668,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2445, "finished": true, "Ev Idx": 4667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564848453.551, "dur": 26.071,
    "args": {
      "External id": 4669,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2446, "finished": true, "Ev Idx": 4668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564848463.274, "dur": 1.024,
    "args": {
      "External id": 4670,"Record function id": 0, "finished": true, "Ev Idx": 4669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564848480.151, "dur": 1.041,
    "args": {
      "External id": 4671,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2447, "finished": true, "Ev Idx": 4670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564848486.010, "dur": 9.088,
    "args": {
      "External id": 4672,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2448, "finished": true, "Ev Idx": 4671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564848504.224, "dur": 0.181,
    "args": {
      "External id": 4673,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2449, "finished": true, "Ev Idx": 4672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564848506.324, "dur": 10.517,
    "args": {
      "External id": 4674,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2449, "finished": true, "Ev Idx": 4673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564848507.955, "dur": 0.331,
    "args": {
      "External id": 4675,"Record function id": 0, "finished": true, "Ev Idx": 4674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564848508.793, "dur": 0.067,
    "args": {
      "External id": 4676,"Record function id": 0, "finished": true, "Ev Idx": 4675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564848519.633, "dur": 13.747,
    "args": {
      "External id": 4677,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2450, "finished": true, "Ev Idx": 4676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564848536.997, "dur": 9.834,
    "args": {
      "External id": 4678,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2451, "finished": true, "Ev Idx": 4677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564848548.810, "dur": 10.396,
    "args": {
      "External id": 4679,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2452, "finished": true, "Ev Idx": 4678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564848560.873, "dur": 7.402,
    "args": {
      "External id": 4680,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2453, "finished": true, "Ev Idx": 4679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564848570.715, "dur": 0.149,
    "args": {
      "External id": 4681,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2454, "finished": true, "Ev Idx": 4680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564848571.716, "dur": 8.852,
    "args": {
      "External id": 4682,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2454, "finished": true, "Ev Idx": 4681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564848595.621, "dur": 108.228,
    "args": {
      "External id": 4683,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2455, "finished": true, "Ev Idx": 4682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564848596.022, "dur": 2.786,
    "args": {
      "External id": 4684,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2455, "finished": true, "Ev Idx": 4683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564848596.956, "dur": 1.297,
    "args": {
      "External id": 4685,"Record function id": 0, "finished": true, "Ev Idx": 4684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848597.671, "dur": 0.407,
    "args": {
      "External id": 4686,"Record function id": 0, "finished": true, "Ev Idx": 4685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564848599.163, "dur": 104.528,
    "args": {
      "External id": 4687,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2456, "finished": true, "Ev Idx": 4686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848599.624, "dur": 2.502,
    "args": {
      "External id": 4688,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2456, "finished": true, "Ev Idx": 4687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564848599.897, "dur": 2.104,
    "args": {
      "External id": 4689,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2456, "finished": true, "Ev Idx": 4688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564848602.445, "dur": 98.727,
    "args": {
      "External id": 4690,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2457, "finished": true, "Ev Idx": 4689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564848672.158, "dur": 1.244,
    "args": {
      "External id": 4691,"Record function id": 0, "finished": true, "Ev Idx": 4690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564848702.160, "dur": 1.174,
    "args": {
      "External id": 4692,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2458, "finished": true, "Ev Idx": 4691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564848713.235, "dur": 10.486,
    "args": {
      "External id": 4693,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2459, "finished": true, "Ev Idx": 4692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564848731.596, "dur": 77.898,
    "args": {
      "External id": 4694,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2460, "finished": true, "Ev Idx": 4693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564848731.943, "dur": 2.830,
    "args": {
      "External id": 4695,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2460, "finished": true, "Ev Idx": 4694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564848732.977, "dur": 1.212,
    "args": {
      "External id": 4696,"Record function id": 0, "finished": true, "Ev Idx": 4695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848733.611, "dur": 0.437,
    "args": {
      "External id": 4697,"Record function id": 0, "finished": true, "Ev Idx": 4696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564848735.118, "dur": 74.281,
    "args": {
      "External id": 4698,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2461, "finished": true, "Ev Idx": 4697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848735.562, "dur": 2.039,
    "args": {
      "External id": 4699,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2461, "finished": true, "Ev Idx": 4698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564848735.920, "dur": 1.555,
    "args": {
      "External id": 4700,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2461, "finished": true, "Ev Idx": 4699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564848737.969, "dur": 69.200,
    "args": {
      "External id": 4701,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2462, "finished": true, "Ev Idx": 4700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564848781.513, "dur": 1.181,
    "args": {
      "External id": 4702,"Record function id": 0, "finished": true, "Ev Idx": 4701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564848807.839, "dur": 1.221,
    "args": {
      "External id": 4703,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2463, "finished": true, "Ev Idx": 4702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564848811.799, "dur": 9.073,
    "args": {
      "External id": 4704,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2464, "finished": true, "Ev Idx": 4703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564848828.694, "dur": 41.375,
    "args": {
      "External id": 4705,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2465, "finished": true, "Ev Idx": 4704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564848829.389, "dur": 3.012,
    "args": {
      "External id": 4706,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2465, "finished": true, "Ev Idx": 4705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564848830.654, "dur": 1.023,
    "args": {
      "External id": 4707,"Record function id": 0, "finished": true, "Ev Idx": 4706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848831.064, "dur": 0.459,
    "args": {
      "External id": 4708,"Record function id": 0, "finished": true, "Ev Idx": 4707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564848832.691, "dur": 37.252,
    "args": {
      "External id": 4709,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2466, "finished": true, "Ev Idx": 4708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564848833.122, "dur": 1.807,
    "args": {
      "External id": 4710,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2466, "finished": true, "Ev Idx": 4709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564848833.437, "dur": 1.381,
    "args": {
      "External id": 4711,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2466, "finished": true, "Ev Idx": 4710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564848835.285, "dur": 32.806,
    "args": {
      "External id": 4712,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2467, "finished": true, "Ev Idx": 4711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564848847.139, "dur": 1.166,
    "args": {
      "External id": 4713,"Record function id": 0, "finished": true, "Ev Idx": 4712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564848868.630, "dur": 1.018,
    "args": {
      "External id": 4714,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2468, "finished": true, "Ev Idx": 4713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564848872.892, "dur": 9.523,
    "args": {
      "External id": 4715,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2469, "finished": true, "Ev Idx": 4714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564848903.308, "dur": 0.208,
    "args": {
      "External id": 4716,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2470, "finished": true, "Ev Idx": 4715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564848905.549, "dur": 11.304,
    "args": {
      "External id": 4717,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2470, "finished": true, "Ev Idx": 4716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564848906.841, "dur": 0.847,
    "args": {
      "External id": 4718,"Record function id": 0, "finished": true, "Ev Idx": 4717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564848908.033, "dur": 0.075,
    "args": {
      "External id": 4719,"Record function id": 0, "finished": true, "Ev Idx": 4718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564848919.569, "dur": 12.685,
    "args": {
      "External id": 4720,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2471, "finished": true, "Ev Idx": 4719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564848935.767, "dur": 7.637,
    "args": {
      "External id": 4721,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2472, "finished": true, "Ev Idx": 4720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564848945.286, "dur": 9.035,
    "args": {
      "External id": 4722,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2473, "finished": true, "Ev Idx": 4721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564848955.988, "dur": 8.126,
    "args": {
      "External id": 4723,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2474, "finished": true, "Ev Idx": 4722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564848966.613, "dur": 0.302,
    "args": {
      "External id": 4724,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2475, "finished": true, "Ev Idx": 4723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564848967.925, "dur": 10.043,
    "args": {
      "External id": 4725,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2475, "finished": true, "Ev Idx": 4724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564848996.447, "dur": 40.343,
    "args": {
      "External id": 4726,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2476, "finished": true, "Ev Idx": 4725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564848997.136, "dur": 3.859,
    "args": {
      "External id": 4727,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2476, "finished": true, "Ev Idx": 4726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564848999.013, "dur": 1.383,
    "args": {
      "External id": 4728,"Record function id": 0, "finished": true, "Ev Idx": 4727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564848999.765, "dur": 0.489,
    "args": {
      "External id": 4729,"Record function id": 0, "finished": true, "Ev Idx": 4728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564849001.395, "dur": 35.312,
    "args": {
      "External id": 4730,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2477, "finished": true, "Ev Idx": 4729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849001.842, "dur": 1.667,
    "args": {
      "External id": 4731,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2477, "finished": true, "Ev Idx": 4730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849002.136, "dur": 1.228,
    "args": {
      "External id": 4732,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2477, "finished": true, "Ev Idx": 4731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564849003.851, "dur": 30.646,
    "args": {
      "External id": 4733,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2478, "finished": true, "Ev Idx": 4732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849016.975, "dur": 1.288,
    "args": {
      "External id": 4734,"Record function id": 0, "finished": true, "Ev Idx": 4733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564849035.225, "dur": 1.154,
    "args": {
      "External id": 4735,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2479, "finished": true, "Ev Idx": 4734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849039.872, "dur": 1.425,
    "args": {
      "External id": 4736,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2480, "finished": true, "Ev Idx": 4735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564849043.401, "dur": 2.848,
    "args": {
      "External id": 4737,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2481, "finished": true, "Ev Idx": 4736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849044.733, "dur": 0.690,
    "args": {
      "External id": 4738,"Record function id": 0, "finished": true, "Ev Idx": 4737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564849053.306, "dur": 43.788,
    "args": {
      "External id": 4739,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2482, "finished": true, "Ev Idx": 4738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564849053.611, "dur": 2.233,
    "args": {
      "External id": 4740,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2482, "finished": true, "Ev Idx": 4739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564849054.546, "dur": 0.842,
    "args": {
      "External id": 4741,"Record function id": 0, "finished": true, "Ev Idx": 4740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849055.013, "dur": 0.264,
    "args": {
      "External id": 4742,"Record function id": 0, "finished": true, "Ev Idx": 4741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564849056.358, "dur": 40.633,
    "args": {
      "External id": 4743,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2483, "finished": true, "Ev Idx": 4742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849056.929, "dur": 1.293,
    "args": {
      "External id": 4744,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2483, "finished": true, "Ev Idx": 4743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849057.214, "dur": 0.888,
    "args": {
      "External id": 4745,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2483, "finished": true, "Ev Idx": 4744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564849058.555, "dur": 36.612,
    "args": {
      "External id": 4746,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2484, "finished": true, "Ev Idx": 4745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849072.212, "dur": 1.023,
    "args": {
      "External id": 4747,"Record function id": 0, "finished": true, "Ev Idx": 4746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564849095.691, "dur": 1.019,
    "args": {
      "External id": 4748,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2485, "finished": true, "Ev Idx": 4747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849099.292, "dur": 1.233,
    "args": {
      "External id": 4749,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2486, "finished": true, "Ev Idx": 4748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564849101.974, "dur": 2.166,
    "args": {
      "External id": 4750,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2487, "finished": true, "Ev Idx": 4749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849102.955, "dur": 0.604,
    "args": {
      "External id": 4751,"Record function id": 0, "finished": true, "Ev Idx": 4750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564849110.476, "dur": 38.209,
    "args": {
      "External id": 4752,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2488, "finished": true, "Ev Idx": 4751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564849110.804, "dur": 1.956,
    "args": {
      "External id": 4753,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2488, "finished": true, "Ev Idx": 4752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564849111.561, "dur": 0.816,
    "args": {
      "External id": 4754,"Record function id": 0, "finished": true, "Ev Idx": 4753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849111.995, "dur": 0.288,
    "args": {
      "External id": 4755,"Record function id": 0, "finished": true, "Ev Idx": 4754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564849113.067, "dur": 35.504,
    "args": {
      "External id": 4756,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2489, "finished": true, "Ev Idx": 4755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849113.409, "dur": 1.838,
    "args": {
      "External id": 4757,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2489, "finished": true, "Ev Idx": 4756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849113.669, "dur": 1.479,
    "args": {
      "External id": 4758,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2489, "finished": true, "Ev Idx": 4757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564849115.574, "dur": 31.326,
    "args": {
      "External id": 4759,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2490, "finished": true, "Ev Idx": 4758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849127.084, "dur": 1.048,
    "args": {
      "External id": 4760,"Record function id": 0, "finished": true, "Ev Idx": 4759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564849147.392, "dur": 0.906,
    "args": {
      "External id": 4761,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2491, "finished": true, "Ev Idx": 4760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849150.892, "dur": 1.207,
    "args": {
      "External id": 4762,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2492, "finished": true, "Ev Idx": 4761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564849153.470, "dur": 1.736,
    "args": {
      "External id": 4763,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2493, "finished": true, "Ev Idx": 4762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849154.271, "dur": 0.449,
    "args": {
      "External id": 4764,"Record function id": 0, "finished": true, "Ev Idx": 4763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564849158.194, "dur": 2.267,
    "args": {
      "External id": 4765,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2494, "finished": true, "Ev Idx": 4764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849159.417, "dur": 0.527,
    "args": {
      "External id": 4766,"Record function id": 0, "finished": true, "Ev Idx": 4765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564849161.613, "dur": 1.199,
    "args": {
      "External id": 4767,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2494, "finished": true, "Ev Idx": 4766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849162.176, "dur": 0.310,
    "args": {
      "External id": 4768,"Record function id": 0, "finished": true, "Ev Idx": 4767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564849163.709, "dur": 10.513,
    "args": {
      "External id": 4769,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2494, "finished": true, "Ev Idx": 4768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849179.580, "dur": 5.307,
    "args": {
      "External id": 4770,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2495, "finished": true, "Ev Idx": 4769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849182.075, "dur": 0.871,
    "args": {
      "External id": 4771,"Record function id": 0, "finished": true, "Ev Idx": 4770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849187.210, "dur": 2.633,
    "args": {
      "External id": 4772,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2496, "finished": true, "Ev Idx": 4771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849188.915, "dur": 0.336,
    "args": {
      "External id": 4773,"Record function id": 0, "finished": true, "Ev Idx": 4772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564849191.640, "dur": 12.636,
    "args": {
      "External id": 4774,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2497, "finished": true, "Ev Idx": 4773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564849208.562, "dur": 18.326,
    "args": {
      "External id": 4775,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2498, "finished": true, "Ev Idx": 4774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564849230.134, "dur": 50.046,
    "args": {
      "External id": 4776,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2499, "finished": true, "Ev Idx": 4775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564849282.549, "dur": 10.869,
    "args": {
      "External id": 4777,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2500, "finished": true, "Ev Idx": 4776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564849295.636, "dur": 8.279,
    "args": {
      "External id": 4778,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2501, "finished": true, "Ev Idx": 4777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849307.286, "dur": 3.482,
    "args": {
      "External id": 4779,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2502, "finished": true, "Ev Idx": 4778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849309.307, "dur": 0.583,
    "args": {
      "External id": 4780,"Record function id": 0, "finished": true, "Ev Idx": 4779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849312.360, "dur": 1.703,
    "args": {
      "External id": 4781,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2503, "finished": true, "Ev Idx": 4780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849313.178, "dur": 0.355,
    "args": {
      "External id": 4782,"Record function id": 0, "finished": true, "Ev Idx": 4781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564849315.219, "dur": 6.927,
    "args": {
      "External id": 4783,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2504, "finished": true, "Ev Idx": 4782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564849324.494, "dur": 12.852,
    "args": {
      "External id": 4784,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2505, "finished": true, "Ev Idx": 4783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564849339.788, "dur": 6.820,
    "args": {
      "External id": 4785,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2506, "finished": true, "Ev Idx": 4784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564849347.999, "dur": 7.569,
    "args": {
      "External id": 4786,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2507, "finished": true, "Ev Idx": 4785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849380.567, "dur": 3.809,
    "args": {
      "External id": 4787,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2508, "finished": true, "Ev Idx": 4786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849382.901, "dur": 0.676,
    "args": {
      "External id": 4788,"Record function id": 0, "finished": true, "Ev Idx": 4787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849385.299, "dur": 2.925,
    "args": {
      "External id": 4789,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2509, "finished": true, "Ev Idx": 4788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849387.267, "dur": 0.491,
    "args": {
      "External id": 4790,"Record function id": 0, "finished": true, "Ev Idx": 4789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564849389.063, "dur": 2.777,
    "args": {
      "External id": 4791,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2510, "finished": true, "Ev Idx": 4790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849390.189, "dur": 0.988,
    "args": {
      "External id": 4792,"Record function id": 0, "finished": true, "Ev Idx": 4791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849392.537, "dur": 1.775,
    "args": {
      "External id": 4793,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2511, "finished": true, "Ev Idx": 4792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849393.408, "dur": 0.403,
    "args": {
      "External id": 4794,"Record function id": 0, "finished": true, "Ev Idx": 4793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849394.927, "dur": 1.470,
    "args": {
      "External id": 4795,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2512, "finished": true, "Ev Idx": 4794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849395.640, "dur": 0.310,
    "args": {
      "External id": 4796,"Record function id": 0, "finished": true, "Ev Idx": 4795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564849399.443, "dur": 4.276,
    "args": {
      "External id": 4797,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2513, "finished": true, "Ev Idx": 4796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849402.569, "dur": 0.421,
    "args": {
      "External id": 4798,"Record function id": 0, "finished": true, "Ev Idx": 4797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849406.171, "dur": 20.459,
    "args": {
      "External id": 4799,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2514, "finished": true, "Ev Idx": 4798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564849408.608, "dur": 15.082,
    "args": {
      "External id": 4800,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2514, "finished": true, "Ev Idx": 4799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564849410.453, "dur": 2.876,
    "args": {
      "External id": 4801,"Record function id": 0, "finished": true, "Ev Idx": 4800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849411.382, "dur": 1.708,
    "args": {
      "External id": 4802,"Record function id": 0, "finished": true, "Ev Idx": 4801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564849413.966, "dur": 9.135,
    "args": {
      "External id": 4803,"Record function id": 0, "finished": true, "Ev Idx": 4802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564849424.251, "dur": 2.116,
    "args": {
      "External id": 4804,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2515, "finished": true, "Ev Idx": 4803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849429.878, "dur": 2.652,
    "args": {
      "External id": 4805,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2516, "finished": true, "Ev Idx": 4804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849431.522, "dur": 0.415,
    "args": {
      "External id": 4806,"Record function id": 0, "finished": true, "Ev Idx": 4805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849433.141, "dur": 1.421,
    "args": {
      "External id": 4807,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2517, "finished": true, "Ev Idx": 4806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849433.831, "dur": 0.300,
    "args": {
      "External id": 4808,"Record function id": 0, "finished": true, "Ev Idx": 4807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564849435.102, "dur": 1.712,
    "args": {
      "External id": 4809,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2518, "finished": true, "Ev Idx": 4808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849435.963, "dur": 0.439,
    "args": {
      "External id": 4810,"Record function id": 0, "finished": true, "Ev Idx": 4809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849437.420, "dur": 2.658,
    "args": {
      "External id": 4811,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2519, "finished": true, "Ev Idx": 4810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849439.293, "dur": 0.336,
    "args": {
      "External id": 4812,"Record function id": 0, "finished": true, "Ev Idx": 4811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849440.805, "dur": 2.330,
    "args": {
      "External id": 4813,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2520, "finished": true, "Ev Idx": 4812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849442.374, "dur": 0.311,
    "args": {
      "External id": 4814,"Record function id": 0, "finished": true, "Ev Idx": 4813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564849445.102, "dur": 1.790,
    "args": {
      "External id": 4815,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2521, "finished": true, "Ev Idx": 4814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849446.101, "dur": 0.365,
    "args": {
      "External id": 4816,"Record function id": 0, "finished": true, "Ev Idx": 4815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849448.452, "dur": 12.077,
    "args": {
      "External id": 4817,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2522, "finished": true, "Ev Idx": 4816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564849449.588, "dur": 9.262,
    "args": {
      "External id": 4818,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2522, "finished": true, "Ev Idx": 4817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564849450.432, "dur": 1.989,
    "args": {
      "External id": 4819,"Record function id": 0, "finished": true, "Ev Idx": 4818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849450.812, "dur": 1.480,
    "args": {
      "External id": 4820,"Record function id": 0, "finished": true, "Ev Idx": 4819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564849452.743, "dur": 5.755,
    "args": {
      "External id": 4821,"Record function id": 0, "finished": true, "Ev Idx": 4820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564849459.224, "dur": 1.123,
    "args": {
      "External id": 4822,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2523, "finished": true, "Ev Idx": 4821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564849462.954, "dur": 2.714,
    "args": {
      "External id": 4823,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2524, "finished": true, "Ev Idx": 4822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849464.702, "dur": 0.418,
    "args": {
      "External id": 4824,"Record function id": 0, "finished": true, "Ev Idx": 4823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564849467.270, "dur": 47.280,
    "args": {
      "External id": 4825,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2525, "finished": true, "Ev Idx": 4824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564849467.878, "dur": 2.425,
    "args": {
      "External id": 4826,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2525, "finished": true, "Ev Idx": 4825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849469.517, "dur": 0.315,
    "args": {
      "External id": 4827,"Record function id": 0, "finished": true, "Ev Idx": 4826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849470.687, "dur": 3.834,
    "args": {
      "External id": 4828,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2526, "finished": true, "Ev Idx": 4827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564849472.485, "dur": 1.865,
    "args": {
      "External id": 4829,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2526, "finished": true, "Ev Idx": 4828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564849475.053, "dur": 1.377,
    "args": {
      "External id": 4830,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2527, "finished": true, "Ev Idx": 4829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849475.781, "dur": 0.274,
    "args": {
      "External id": 4831,"Record function id": 0, "finished": true, "Ev Idx": 4830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849476.737, "dur": 1.824,
    "args": {
      "External id": 4832,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2528, "finished": true, "Ev Idx": 4831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564849477.567, "dur": 0.894,
    "args": {
      "External id": 4833,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2528, "finished": true, "Ev Idx": 4832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564849479.216, "dur": 33.248,
    "args": {
      "External id": 4834,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2529, "finished": true, "Ev Idx": 4833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849492.985, "dur": 0.990,
    "args": {
      "External id": 4835,"Record function id": 0, "finished": true, "Ev Idx": 4834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564849513.049, "dur": 1.153,
    "args": {
      "External id": 4836,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2530, "finished": true, "Ev Idx": 4835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564849517.878, "dur": 9.494,
    "args": {
      "External id": 4837,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2531, "finished": true, "Ev Idx": 4836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849530.297, "dur": 1.967,
    "args": {
      "External id": 4838,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2532, "finished": true, "Ev Idx": 4837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849531.089, "dur": 0.489,
    "args": {
      "External id": 4839,"Record function id": 0, "finished": true, "Ev Idx": 4838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849532.843, "dur": 1.203,
    "args": {
      "External id": 4840,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2532, "finished": true, "Ev Idx": 4839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849533.364, "dur": 0.337,
    "args": {
      "External id": 4841,"Record function id": 0, "finished": true, "Ev Idx": 4840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564849534.681, "dur": 1.024,
    "args": {
      "External id": 4842,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2532, "finished": true, "Ev Idx": 4841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849535.131, "dur": 0.262,
    "args": {
      "External id": 4843,"Record function id": 0, "finished": true, "Ev Idx": 4842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564849537.609, "dur": 7.836,
    "args": {
      "External id": 4844,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2532, "finished": true, "Ev Idx": 4843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564849550.988, "dur": 13.020,
    "args": {
      "External id": 4845,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2533, "finished": true, "Ev Idx": 4844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564849551.924, "dur": 0.427,
    "args": {
      "External id": 4846,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2533, "finished": true, "Ev Idx": 4845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564849552.985, "dur": 10.719,
    "args": {
      "External id": 4847,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2533, "finished": true, "Ev Idx": 4846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564849566.959, "dur": 0.285,
    "args": {
      "External id": 4848,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2534, "finished": true, "Ev Idx": 4847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564849578.546, "dur": 0.446,
    "args": {
      "External id": 4849,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2534, "finished": true, "Ev Idx": 4848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564849580.945, "dur": 40.898,
    "args": {
      "External id": 4850,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2534, "finished": true, "Ev Idx": 4849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564849581.616, "dur": 2.279,
    "args": {
      "External id": 4851,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2534, "finished": true, "Ev Idx": 4850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849582.872, "dur": 0.450,
    "args": {
      "External id": 4852,"Record function id": 0, "finished": true, "Ev Idx": 4851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849584.235, "dur": 4.196,
    "args": {
      "External id": 4853,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2535, "finished": true, "Ev Idx": 4852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849584.583, "dur": 3.704,
    "args": {
      "External id": 4854,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2535, "finished": true, "Ev Idx": 4853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564849588.904, "dur": 2.207,
    "args": {
      "External id": 4855,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2536, "finished": true, "Ev Idx": 4854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849590.502, "dur": 0.228,
    "args": {
      "External id": 4856,"Record function id": 0, "finished": true, "Ev Idx": 4855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849591.410, "dur": 1.194,
    "args": {
      "External id": 4857,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2537, "finished": true, "Ev Idx": 4856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849591.702, "dur": 0.832,
    "args": {
      "External id": 4858,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2537, "finished": true, "Ev Idx": 4857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564849593.011, "dur": 26.848,
    "args": {
      "External id": 4859,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2538, "finished": true, "Ev Idx": 4858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849603.488, "dur": 1.068,
    "args": {
      "External id": 4860,"Record function id": 0, "finished": true, "Ev Idx": 4859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564849620.431, "dur": 1.081,
    "args": {
      "External id": 4861,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2539, "finished": true, "Ev Idx": 4860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564849623.958, "dur": 2.038,
    "args": {
      "External id": 4862,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2540, "finished": true, "Ev Idx": 4861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849624.968, "dur": 0.518,
    "args": {
      "External id": 4863,"Record function id": 0, "finished": true, "Ev Idx": 4862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564849628.488, "dur": 12.569,
    "args": {
      "External id": 4864,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2541, "finished": true, "Ev Idx": 4863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564849629.489, "dur": 11.430,
    "args": {
      "External id": 4865,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2541, "finished": true, "Ev Idx": 4864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564849630.687, "dur": 1.987,
    "args": {
      "External id": 4866,"Record function id": 0, "finished": true, "Ev Idx": 4865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849631.063, "dur": 1.451,
    "args": {
      "External id": 4867,"Record function id": 0, "finished": true, "Ev Idx": 4866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564849633.029, "dur": 7.539,
    "args": {
      "External id": 4868,"Record function id": 0, "finished": true, "Ev Idx": 4867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849646.771, "dur": 1.885,
    "args": {
      "External id": 4869,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2542, "finished": true, "Ev Idx": 4868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849647.318, "dur": 1.237,
    "args": {
      "External id": 4870,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2542, "finished": true, "Ev Idx": 4869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564849663.021, "dur": 36.310,
    "args": {
      "External id": 4871,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2543, "finished": true, "Ev Idx": 4870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564849663.742, "dur": 3.037,
    "args": {
      "External id": 4872,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2543, "finished": true, "Ev Idx": 4871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564849665.131, "dur": 1.092,
    "args": {
      "External id": 4873,"Record function id": 0, "finished": true, "Ev Idx": 4872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849665.673, "dur": 0.392,
    "args": {
      "External id": 4874,"Record function id": 0, "finished": true, "Ev Idx": 4873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564849667.175, "dur": 32.018,
    "args": {
      "External id": 4875,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2544, "finished": true, "Ev Idx": 4874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849667.749, "dur": 1.775,
    "args": {
      "External id": 4876,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2544, "finished": true, "Ev Idx": 4875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849668.204, "dur": 1.219,
    "args": {
      "External id": 4877,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2544, "finished": true, "Ev Idx": 4876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564849669.955, "dur": 27.152,
    "args": {
      "External id": 4878,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2545, "finished": true, "Ev Idx": 4877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849681.294, "dur": 1.127,
    "args": {
      "External id": 4879,"Record function id": 0, "finished": true, "Ev Idx": 4878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564849697.732, "dur": 1.133,
    "args": {
      "External id": 4880,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2546, "finished": true, "Ev Idx": 4879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564849704.161, "dur": 8.836,
    "args": {
      "External id": 4881,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2547, "finished": true, "Ev Idx": 4880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564849721.712, "dur": 0.192,
    "args": {
      "External id": 4882,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2548, "finished": true, "Ev Idx": 4881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564849724.023, "dur": 10.408,
    "args": {
      "External id": 4883,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2548, "finished": true, "Ev Idx": 4882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564849725.724, "dur": 0.362,
    "args": {
      "External id": 4884,"Record function id": 0, "finished": true, "Ev Idx": 4883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564849726.426, "dur": 0.081,
    "args": {
      "External id": 4885,"Record function id": 0, "finished": true, "Ev Idx": 4884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564849737.459, "dur": 11.894,
    "args": {
      "External id": 4886,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2549, "finished": true, "Ev Idx": 4885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564849752.987, "dur": 9.394,
    "args": {
      "External id": 4887,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2550, "finished": true, "Ev Idx": 4886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564849764.403, "dur": 9.994,
    "args": {
      "External id": 4888,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2551, "finished": true, "Ev Idx": 4887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564849776.154, "dur": 9.115,
    "args": {
      "External id": 4889,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2552, "finished": true, "Ev Idx": 4888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564849787.982, "dur": 0.098,
    "args": {
      "External id": 4890,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2553, "finished": true, "Ev Idx": 4889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564849789.115, "dur": 10.343,
    "args": {
      "External id": 4891,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2553, "finished": true, "Ev Idx": 4890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564849812.913, "dur": 49.253,
    "args": {
      "External id": 4892,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2554, "finished": true, "Ev Idx": 4891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564849813.601, "dur": 3.643,
    "args": {
      "External id": 4893,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2554, "finished": true, "Ev Idx": 4892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564849815.023, "dur": 1.664,
    "args": {
      "External id": 4894,"Record function id": 0, "finished": true, "Ev Idx": 4893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849815.913, "dur": 0.613,
    "args": {
      "External id": 4895,"Record function id": 0, "finished": true, "Ev Idx": 4894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564849817.598, "dur": 44.437,
    "args": {
      "External id": 4896,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2555, "finished": true, "Ev Idx": 4895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849818.157, "dur": 3.160,
    "args": {
      "External id": 4897,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2555, "finished": true, "Ev Idx": 4896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849818.596, "dur": 2.611,
    "args": {
      "External id": 4898,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2555, "finished": true, "Ev Idx": 4897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564849821.800, "dur": 38.377,
    "args": {
      "External id": 4899,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2556, "finished": true, "Ev Idx": 4898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849833.361, "dur": 1.013,
    "args": {
      "External id": 4900,"Record function id": 0, "finished": true, "Ev Idx": 4899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564849860.760, "dur": 0.966,
    "args": {
      "External id": 4901,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2557, "finished": true, "Ev Idx": 4900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564849870.269, "dur": 59.040,
    "args": {
      "External id": 4902,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2558, "finished": true, "Ev Idx": 4901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564849939.985, "dur": 46.859,
    "args": {
      "External id": 4903,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2559, "finished": true, "Ev Idx": 4902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564849940.424, "dur": 3.633,
    "args": {
      "External id": 4904,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2559, "finished": true, "Ev Idx": 4903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564849942.115, "dur": 1.174,
    "args": {
      "External id": 4905,"Record function id": 0, "finished": true, "Ev Idx": 4904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564849942.622, "dur": 0.551,
    "args": {
      "External id": 4906,"Record function id": 0, "finished": true, "Ev Idx": 4905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564849944.390, "dur": 42.348,
    "args": {
      "External id": 4907,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2560, "finished": true, "Ev Idx": 4906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564849944.904, "dur": 1.669,
    "args": {
      "External id": 4908,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2560, "finished": true, "Ev Idx": 4907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564849945.225, "dur": 1.247,
    "args": {
      "External id": 4909,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2560, "finished": true, "Ev Idx": 4908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564849946.893, "dur": 37.991,
    "args": {
      "External id": 4910,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2561, "finished": true, "Ev Idx": 4909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564849960.423, "dur": 1.171,
    "args": {
      "External id": 4911,"Record function id": 0, "finished": true, "Ev Idx": 4910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564849985.448, "dur": 0.946,
    "args": {
      "External id": 4912,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2562, "finished": true, "Ev Idx": 4911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564849988.753, "dur": 41.791,
    "args": {
      "External id": 4913,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2563, "finished": true, "Ev Idx": 4912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564850038.172, "dur": 61.292,
    "args": {
      "External id": 4914,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2564, "finished": true, "Ev Idx": 4913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564850038.504, "dur": 3.469,
    "args": {
      "External id": 4915,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2564, "finished": true, "Ev Idx": 4914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850039.800, "dur": 1.292,
    "args": {
      "External id": 4916,"Record function id": 0, "finished": true, "Ev Idx": 4915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850040.324, "dur": 0.637,
    "args": {
      "External id": 4917,"Record function id": 0, "finished": true, "Ev Idx": 4916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564850042.274, "dur": 57.082,
    "args": {
      "External id": 4918,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2565, "finished": true, "Ev Idx": 4917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850042.779, "dur": 1.618,
    "args": {
      "External id": 4919,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2565, "finished": true, "Ev Idx": 4918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850043.071, "dur": 1.233,
    "args": {
      "External id": 4920,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2565, "finished": true, "Ev Idx": 4919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564850044.722, "dur": 50.989,
    "args": {
      "External id": 4921,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2566, "finished": true, "Ev Idx": 4920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564850054.910, "dur": 1.073,
    "args": {
      "External id": 4922,"Record function id": 0, "finished": true, "Ev Idx": 4921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564850097.257, "dur": 1.644,
    "args": {
      "External id": 4923,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2567, "finished": true, "Ev Idx": 4922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564850104.167, "dur": 11.785,
    "args": {
      "External id": 4924,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2568, "finished": true, "Ev Idx": 4923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564850136.917, "dur": 0.215,
    "args": {
      "External id": 4925,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2569, "finished": true, "Ev Idx": 4924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564850139.572, "dur": 11.668,
    "args": {
      "External id": 4926,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2569, "finished": true, "Ev Idx": 4925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564850141.516, "dur": 0.604,
    "args": {
      "External id": 4927,"Record function id": 0, "finished": true, "Ev Idx": 4926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564850142.759, "dur": 0.089,
    "args": {
      "External id": 4928,"Record function id": 0, "finished": true, "Ev Idx": 4927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564850154.148, "dur": 13.035,
    "args": {
      "External id": 4929,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2570, "finished": true, "Ev Idx": 4928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564850171.303, "dur": 10.237,
    "args": {
      "External id": 4930,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2571, "finished": true, "Ev Idx": 4929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564850183.600, "dur": 9.558,
    "args": {
      "External id": 4931,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2572, "finished": true, "Ev Idx": 4930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564850194.970, "dur": 7.791,
    "args": {
      "External id": 4932,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2573, "finished": true, "Ev Idx": 4931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564850205.370, "dur": 0.111,
    "args": {
      "External id": 4933,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2574, "finished": true, "Ev Idx": 4932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564850206.272, "dur": 8.179,
    "args": {
      "External id": 4934,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2574, "finished": true, "Ev Idx": 4933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564850233.211, "dur": 41.289,
    "args": {
      "External id": 4935,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2575, "finished": true, "Ev Idx": 4934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564850233.719, "dur": 4.997,
    "args": {
      "External id": 4936,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2575, "finished": true, "Ev Idx": 4935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850235.637, "dur": 1.914,
    "args": {
      "External id": 4937,"Record function id": 0, "finished": true, "Ev Idx": 4936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850236.710, "dur": 0.644,
    "args": {
      "External id": 4938,"Record function id": 0, "finished": true, "Ev Idx": 4937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564850239.198, "dur": 35.208,
    "args": {
      "External id": 4939,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2576, "finished": true, "Ev Idx": 4938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850239.737, "dur": 1.809,
    "args": {
      "External id": 4940,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2576, "finished": true, "Ev Idx": 4939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850240.182, "dur": 1.244,
    "args": {
      "External id": 4941,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2576, "finished": true, "Ev Idx": 4940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564850241.993, "dur": 30.492,
    "args": {
      "External id": 4942,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2577, "finished": true, "Ev Idx": 4941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564850255.094, "dur": 1.237,
    "args": {
      "External id": 4943,"Record function id": 0, "finished": true, "Ev Idx": 4942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564850273.062, "dur": 1.048,
    "args": {
      "External id": 4944,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2578, "finished": true, "Ev Idx": 4943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850277.705, "dur": 1.389,
    "args": {
      "External id": 4945,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2579, "finished": true, "Ev Idx": 4944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850280.942, "dur": 2.857,
    "args": {
      "External id": 4946,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2580, "finished": true, "Ev Idx": 4945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850282.368, "dur": 0.617,
    "args": {
      "External id": 4947,"Record function id": 0, "finished": true, "Ev Idx": 4946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564850290.694, "dur": 42.311,
    "args": {
      "External id": 4948,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2581, "finished": true, "Ev Idx": 4947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564850291.039, "dur": 2.654,
    "args": {
      "External id": 4949,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2581, "finished": true, "Ev Idx": 4948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850292.257, "dur": 0.992,
    "args": {
      "External id": 4950,"Record function id": 0, "finished": true, "Ev Idx": 4949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850292.844, "dur": 0.281,
    "args": {
      "External id": 4951,"Record function id": 0, "finished": true, "Ev Idx": 4950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564850294.182, "dur": 38.724,
    "args": {
      "External id": 4952,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2582, "finished": true, "Ev Idx": 4951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850294.688, "dur": 1.424,
    "args": {
      "External id": 4953,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2582, "finished": true, "Ev Idx": 4952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850295.107, "dur": 0.895,
    "args": {
      "External id": 4954,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2582, "finished": true, "Ev Idx": 4953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564850296.407, "dur": 34.528,
    "args": {
      "External id": 4955,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2583, "finished": true, "Ev Idx": 4954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564850306.784, "dur": 1.125,
    "args": {
      "External id": 4956,"Record function id": 0, "finished": true, "Ev Idx": 4955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564850331.505, "dur": 1.091,
    "args": {
      "External id": 4957,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2584, "finished": true, "Ev Idx": 4956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850335.192, "dur": 1.337,
    "args": {
      "External id": 4958,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2585, "finished": true, "Ev Idx": 4957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850337.935, "dur": 1.645,
    "args": {
      "External id": 4959,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2586, "finished": true, "Ev Idx": 4958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850338.681, "dur": 0.376,
    "args": {
      "External id": 4960,"Record function id": 0, "finished": true, "Ev Idx": 4959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564850346.346, "dur": 37.485,
    "args": {
      "External id": 4961,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2587, "finished": true, "Ev Idx": 4960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564850346.688, "dur": 2.039,
    "args": {
      "External id": 4962,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2587, "finished": true, "Ev Idx": 4961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850347.521, "dur": 0.771,
    "args": {
      "External id": 4963,"Record function id": 0, "finished": true, "Ev Idx": 4962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850347.906, "dur": 0.283,
    "args": {
      "External id": 4964,"Record function id": 0, "finished": true, "Ev Idx": 4963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564850349.021, "dur": 34.713,
    "args": {
      "External id": 4965,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2588, "finished": true, "Ev Idx": 4964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850349.551, "dur": 1.729,
    "args": {
      "External id": 4966,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2588, "finished": true, "Ev Idx": 4965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850350.291, "dur": 0.886,
    "args": {
      "External id": 4967,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2588, "finished": true, "Ev Idx": 4966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564850351.565, "dur": 30.377,
    "args": {
      "External id": 4968,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2589, "finished": true, "Ev Idx": 4967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564850362.349, "dur": 1.037,
    "args": {
      "External id": 4969,"Record function id": 0, "finished": true, "Ev Idx": 4968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564850382.546, "dur": 0.920,
    "args": {
      "External id": 4970,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2590, "finished": true, "Ev Idx": 4969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850385.950, "dur": 1.099,
    "args": {
      "External id": 4971,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2591, "finished": true, "Ev Idx": 4970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850388.504, "dur": 1.531,
    "args": {
      "External id": 4972,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2592, "finished": true, "Ev Idx": 4971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850389.226, "dur": 0.346,
    "args": {
      "External id": 4973,"Record function id": 0, "finished": true, "Ev Idx": 4972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564850392.734, "dur": 2.180,
    "args": {
      "External id": 4974,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2593, "finished": true, "Ev Idx": 4973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850394.086, "dur": 0.318,
    "args": {
      "External id": 4975,"Record function id": 0, "finished": true, "Ev Idx": 4974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564850396.019, "dur": 1.066,
    "args": {
      "External id": 4976,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2593, "finished": true, "Ev Idx": 4975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850396.513, "dur": 0.291,
    "args": {
      "External id": 4977,"Record function id": 0, "finished": true, "Ev Idx": 4976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564850398.301, "dur": 9.476,
    "args": {
      "External id": 4978,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2593, "finished": true, "Ev Idx": 4977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850413.651, "dur": 6.108,
    "args": {
      "External id": 4979,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2594, "finished": true, "Ev Idx": 4978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850416.890, "dur": 1.074,
    "args": {
      "External id": 4980,"Record function id": 0, "finished": true, "Ev Idx": 4979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850422.066, "dur": 1.850,
    "args": {
      "External id": 4981,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2595, "finished": true, "Ev Idx": 4980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850422.971, "dur": 0.351,
    "args": {
      "External id": 4982,"Record function id": 0, "finished": true, "Ev Idx": 4981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564850425.664, "dur": 9.886,
    "args": {
      "External id": 4983,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2596, "finished": true, "Ev Idx": 4982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564850439.593, "dur": 18.476,
    "args": {
      "External id": 4984,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2597, "finished": true, "Ev Idx": 4983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564850460.775, "dur": 8.650,
    "args": {
      "External id": 4985,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2598, "finished": true, "Ev Idx": 4984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564850470.990, "dur": 8.537,
    "args": {
      "External id": 4986,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2599, "finished": true, "Ev Idx": 4985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564850481.409, "dur": 7.540,
    "args": {
      "External id": 4987,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2600, "finished": true, "Ev Idx": 4986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850491.395, "dur": 2.437,
    "args": {
      "External id": 4988,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2601, "finished": true, "Ev Idx": 4987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850492.592, "dur": 0.563,
    "args": {
      "External id": 4989,"Record function id": 0, "finished": true, "Ev Idx": 4988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850495.383, "dur": 1.870,
    "args": {
      "External id": 4990,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2602, "finished": true, "Ev Idx": 4989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850496.388, "dur": 0.337,
    "args": {
      "External id": 4991,"Record function id": 0, "finished": true, "Ev Idx": 4990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564850498.308, "dur": 8.716,
    "args": {
      "External id": 4992,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2603, "finished": true, "Ev Idx": 4991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564850508.760, "dur": 9.470,
    "args": {
      "External id": 4993,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2604, "finished": true, "Ev Idx": 4992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564850520.322, "dur": 6.603,
    "args": {
      "External id": 4994,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2605, "finished": true, "Ev Idx": 4993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564850528.319, "dur": 9.100,
    "args": {
      "External id": 4995,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2606, "finished": true, "Ev Idx": 4994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850561.326, "dur": 2.444,
    "args": {
      "External id": 4996,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2607, "finished": true, "Ev Idx": 4995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850562.609, "dur": 0.450,
    "args": {
      "External id": 4997,"Record function id": 0, "finished": true, "Ev Idx": 4996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850564.430, "dur": 2.598,
    "args": {
      "External id": 4998,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2608, "finished": true, "Ev Idx": 4997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850566.101, "dur": 0.395,
    "args": {
      "External id": 4999,"Record function id": 0, "finished": true, "Ev Idx": 4998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564850567.781, "dur": 2.389,
    "args": {
      "External id": 5000,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2609, "finished": true, "Ev Idx": 4999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850568.835, "dur": 0.707,
    "args": {
      "External id": 5001,"Record function id": 0, "finished": true, "Ev Idx": 5000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850570.861, "dur": 2.227,
    "args": {
      "External id": 5002,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2610, "finished": true, "Ev Idx": 5001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850572.216, "dur": 0.361,
    "args": {
      "External id": 5003,"Record function id": 0, "finished": true, "Ev Idx": 5002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850574.006, "dur": 1.778,
    "args": {
      "External id": 5004,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2611, "finished": true, "Ev Idx": 5003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850575.067, "dur": 0.279,
    "args": {
      "External id": 5005,"Record function id": 0, "finished": true, "Ev Idx": 5004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564850578.756, "dur": 3.957,
    "args": {
      "External id": 5006,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2612, "finished": true, "Ev Idx": 5005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850581.616, "dur": 0.398,
    "args": {
      "External id": 5007,"Record function id": 0, "finished": true, "Ev Idx": 5006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850585.238, "dur": 21.654,
    "args": {
      "External id": 5008,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2613, "finished": true, "Ev Idx": 5007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564850587.858, "dur": 17.128,
    "args": {
      "External id": 5009,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2613, "finished": true, "Ev Idx": 5008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564850589.913, "dur": 2.974,
    "args": {
      "External id": 5010,"Record function id": 0, "finished": true, "Ev Idx": 5009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564850590.998, "dur": 1.610,
    "args": {
      "External id": 5011,"Record function id": 0, "finished": true, "Ev Idx": 5010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564850593.560, "dur": 10.758,
    "args": {
      "External id": 5012,"Record function id": 0, "finished": true, "Ev Idx": 5011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564850605.468, "dur": 1.197,
    "args": {
      "External id": 5013,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2614, "finished": true, "Ev Idx": 5012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850609.994, "dur": 2.759,
    "args": {
      "External id": 5014,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2615, "finished": true, "Ev Idx": 5013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850611.759, "dur": 0.394,
    "args": {
      "External id": 5015,"Record function id": 0, "finished": true, "Ev Idx": 5014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850613.331, "dur": 2.421,
    "args": {
      "External id": 5016,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2616, "finished": true, "Ev Idx": 5015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850614.998, "dur": 0.284,
    "args": {
      "External id": 5017,"Record function id": 0, "finished": true, "Ev Idx": 5016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564850616.263, "dur": 2.721,
    "args": {
      "External id": 5018,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2617, "finished": true, "Ev Idx": 5017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850618.201, "dur": 0.318,
    "args": {
      "External id": 5019,"Record function id": 0, "finished": true, "Ev Idx": 5018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850620.001, "dur": 2.368,
    "args": {
      "External id": 5020,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2618, "finished": true, "Ev Idx": 5019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850621.559, "dur": 0.313,
    "args": {
      "External id": 5021,"Record function id": 0, "finished": true, "Ev Idx": 5020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850623.063, "dur": 2.016,
    "args": {
      "External id": 5022,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2619, "finished": true, "Ev Idx": 5021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850624.267, "dur": 0.394,
    "args": {
      "External id": 5023,"Record function id": 0, "finished": true, "Ev Idx": 5022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564850626.590, "dur": 1.603,
    "args": {
      "External id": 5024,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2620, "finished": true, "Ev Idx": 5023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850627.393, "dur": 0.342,
    "args": {
      "External id": 5025,"Record function id": 0, "finished": true, "Ev Idx": 5024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850629.537, "dur": 12.285,
    "args": {
      "External id": 5026,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2621, "finished": true, "Ev Idx": 5025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564850630.533, "dur": 8.947,
    "args": {
      "External id": 5027,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2621, "finished": true, "Ev Idx": 5026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564850631.311, "dur": 2.049,
    "args": {
      "External id": 5028,"Record function id": 0, "finished": true, "Ev Idx": 5027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564850631.811, "dur": 1.397,
    "args": {
      "External id": 5029,"Record function id": 0, "finished": true, "Ev Idx": 5028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564850633.709, "dur": 5.395,
    "args": {
      "External id": 5030,"Record function id": 0, "finished": true, "Ev Idx": 5029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564850639.990, "dur": 1.686,
    "args": {
      "External id": 5031,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2622, "finished": true, "Ev Idx": 5030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850643.849, "dur": 2.189,
    "args": {
      "External id": 5032,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2623, "finished": true, "Ev Idx": 5031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850645.193, "dur": 0.377,
    "args": {
      "External id": 5033,"Record function id": 0, "finished": true, "Ev Idx": 5032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564850647.576, "dur": 45.441,
    "args": {
      "External id": 5034,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2624, "finished": true, "Ev Idx": 5033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564850648.176, "dur": 1.797,
    "args": {
      "External id": 5035,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2624, "finished": true, "Ev Idx": 5034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850649.238, "dur": 0.331,
    "args": {
      "External id": 5036,"Record function id": 0, "finished": true, "Ev Idx": 5035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850650.303, "dur": 3.900,
    "args": {
      "External id": 5037,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2625, "finished": true, "Ev Idx": 5036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564850652.173, "dur": 1.879,
    "args": {
      "External id": 5038,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2625, "finished": true, "Ev Idx": 5037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564850654.711, "dur": 1.317,
    "args": {
      "External id": 5039,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2626, "finished": true, "Ev Idx": 5038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850655.371, "dur": 0.293,
    "args": {
      "External id": 5040,"Record function id": 0, "finished": true, "Ev Idx": 5039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850656.321, "dur": 1.798,
    "args": {
      "External id": 5041,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2627, "finished": true, "Ev Idx": 5040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564850657.168, "dur": 0.814,
    "args": {
      "External id": 5042,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2627, "finished": true, "Ev Idx": 5041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564850658.763, "dur": 32.209,
    "args": {
      "External id": 5043,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2628, "finished": true, "Ev Idx": 5042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564850673.126, "dur": 1.004,
    "args": {
      "External id": 5044,"Record function id": 0, "finished": true, "Ev Idx": 5043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564850691.548, "dur": 1.145,
    "args": {
      "External id": 5045,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2629, "finished": true, "Ev Idx": 5044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564850696.126, "dur": 10.327,
    "args": {
      "External id": 5046,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2630, "finished": true, "Ev Idx": 5045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850709.443, "dur": 1.877,
    "args": {
      "External id": 5047,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2631, "finished": true, "Ev Idx": 5046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850710.155, "dur": 0.483,
    "args": {
      "External id": 5048,"Record function id": 0, "finished": true, "Ev Idx": 5047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850711.946, "dur": 1.306,
    "args": {
      "External id": 5049,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2631, "finished": true, "Ev Idx": 5048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850712.530, "dur": 0.348,
    "args": {
      "External id": 5050,"Record function id": 0, "finished": true, "Ev Idx": 5049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564850713.863, "dur": 1.048,
    "args": {
      "External id": 5051,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2631, "finished": true, "Ev Idx": 5050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850714.314, "dur": 0.271,
    "args": {
      "External id": 5052,"Record function id": 0, "finished": true, "Ev Idx": 5051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564850716.709, "dur": 8.500,
    "args": {
      "External id": 5053,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2631, "finished": true, "Ev Idx": 5052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564850730.778, "dur": 14.552,
    "args": {
      "External id": 5054,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2632, "finished": true, "Ev Idx": 5053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564850731.996, "dur": 0.370,
    "args": {
      "External id": 5055,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2632, "finished": true, "Ev Idx": 5054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564850733.110, "dur": 11.951,
    "args": {
      "External id": 5056,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2632, "finished": true, "Ev Idx": 5055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564850748.110, "dur": 0.182,
    "args": {
      "External id": 5057,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2633, "finished": true, "Ev Idx": 5056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564850759.745, "dur": 0.478,
    "args": {
      "External id": 5058,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2633, "finished": true, "Ev Idx": 5057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564850762.125, "dur": 42.040,
    "args": {
      "External id": 5059,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2633, "finished": true, "Ev Idx": 5058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564850762.734, "dur": 1.961,
    "args": {
      "External id": 5060,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2633, "finished": true, "Ev Idx": 5059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850763.770, "dur": 0.410,
    "args": {
      "External id": 5061,"Record function id": 0, "finished": true, "Ev Idx": 5060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850765.042, "dur": 2.923,
    "args": {
      "External id": 5062,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2634, "finished": true, "Ev Idx": 5061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850765.414, "dur": 2.411,
    "args": {
      "External id": 5063,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2634, "finished": true, "Ev Idx": 5062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564850768.389, "dur": 2.450,
    "args": {
      "External id": 5064,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2635, "finished": true, "Ev Idx": 5063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850770.066, "dur": 0.383,
    "args": {
      "External id": 5065,"Record function id": 0, "finished": true, "Ev Idx": 5064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850771.128, "dur": 1.707,
    "args": {
      "External id": 5066,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2636, "finished": true, "Ev Idx": 5065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850771.409, "dur": 1.361,
    "args": {
      "External id": 5067,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2636, "finished": true, "Ev Idx": 5066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564850773.238, "dur": 28.858,
    "args": {
      "External id": 5068,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2637, "finished": true, "Ev Idx": 5067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564850783.921, "dur": 1.395,
    "args": {
      "External id": 5069,"Record function id": 0, "finished": true, "Ev Idx": 5068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564850802.652, "dur": 1.208,
    "args": {
      "External id": 5070,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2638, "finished": true, "Ev Idx": 5069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850806.216, "dur": 1.777,
    "args": {
      "External id": 5071,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2639, "finished": true, "Ev Idx": 5070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850807.063, "dur": 0.384,
    "args": {
      "External id": 5072,"Record function id": 0, "finished": true, "Ev Idx": 5071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564850810.508, "dur": 10.789,
    "args": {
      "External id": 5073,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2640, "finished": true, "Ev Idx": 5072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564850811.092, "dur": 10.063,
    "args": {
      "External id": 5074,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2640, "finished": true, "Ev Idx": 5073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564850812.159, "dur": 1.778,
    "args": {
      "External id": 5075,"Record function id": 0, "finished": true, "Ev Idx": 5074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564850812.506, "dur": 1.295,
    "args": {
      "External id": 5076,"Record function id": 0, "finished": true, "Ev Idx": 5075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564850814.221, "dur": 6.484,
    "args": {
      "External id": 5077,"Record function id": 0, "finished": true, "Ev Idx": 5076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850826.773, "dur": 1.747,
    "args": {
      "External id": 5078,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2641, "finished": true, "Ev Idx": 5077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850827.107, "dur": 1.300,
    "args": {
      "External id": 5079,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2641, "finished": true, "Ev Idx": 5078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564850842.706, "dur": 36.289,
    "args": {
      "External id": 5080,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2642, "finished": true, "Ev Idx": 5079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564850843.264, "dur": 3.140,
    "args": {
      "External id": 5081,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2642, "finished": true, "Ev Idx": 5080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850844.579, "dur": 1.257,
    "args": {
      "External id": 5082,"Record function id": 0, "finished": true, "Ev Idx": 5081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850845.331, "dur": 0.383,
    "args": {
      "External id": 5083,"Record function id": 0, "finished": true, "Ev Idx": 5082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564850847.118, "dur": 31.709,
    "args": {
      "External id": 5084,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2643, "finished": true, "Ev Idx": 5083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850847.778, "dur": 1.714,
    "args": {
      "External id": 5085,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2643, "finished": true, "Ev Idx": 5084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850848.388, "dur": 1.002,
    "args": {
      "External id": 5086,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2643, "finished": true, "Ev Idx": 5085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564850850.123, "dur": 26.588,
    "args": {
      "External id": 5087,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2644, "finished": true, "Ev Idx": 5086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564850861.746, "dur": 1.109,
    "args": {
      "External id": 5088,"Record function id": 0, "finished": true, "Ev Idx": 5087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564850877.361, "dur": 1.097,
    "args": {
      "External id": 5089,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2645, "finished": true, "Ev Idx": 5088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564850883.852, "dur": 8.282,
    "args": {
      "External id": 5090,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2646, "finished": true, "Ev Idx": 5089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564850900.706, "dur": 0.189,
    "args": {
      "External id": 5091,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2647, "finished": true, "Ev Idx": 5090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564850902.828, "dur": 11.440,
    "args": {
      "External id": 5092,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2647, "finished": true, "Ev Idx": 5091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564850904.497, "dur": 0.393,
    "args": {
      "External id": 5093,"Record function id": 0, "finished": true, "Ev Idx": 5092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564850905.381, "dur": 0.400,
    "args": {
      "External id": 5094,"Record function id": 0, "finished": true, "Ev Idx": 5093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564850917.084, "dur": 11.808,
    "args": {
      "External id": 5095,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2648, "finished": true, "Ev Idx": 5094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564850932.448, "dur": 8.166,
    "args": {
      "External id": 5096,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2649, "finished": true, "Ev Idx": 5095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564850942.510, "dur": 9.347,
    "args": {
      "External id": 5097,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2650, "finished": true, "Ev Idx": 5096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564850953.772, "dur": 7.415,
    "args": {
      "External id": 5098,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2651, "finished": true, "Ev Idx": 5097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564850963.818, "dur": 0.237,
    "args": {
      "External id": 5099,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2652, "finished": true, "Ev Idx": 5098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564850964.945, "dur": 8.534,
    "args": {
      "External id": 5100,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2652, "finished": true, "Ev Idx": 5099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564850987.990, "dur": 118.999,
    "args": {
      "External id": 5101,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2653, "finished": true, "Ev Idx": 5100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564850988.629, "dur": 3.150,
    "args": {
      "External id": 5102,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2653, "finished": true, "Ev Idx": 5101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564850989.734, "dur": 1.473,
    "args": {
      "External id": 5103,"Record function id": 0, "finished": true, "Ev Idx": 5102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564850990.556, "dur": 0.558,
    "args": {
      "External id": 5104,"Record function id": 0, "finished": true, "Ev Idx": 5103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564850992.112, "dur": 114.737,
    "args": {
      "External id": 5105,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2654, "finished": true, "Ev Idx": 5104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564850992.894, "dur": 2.637,
    "args": {
      "External id": 5106,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2654, "finished": true, "Ev Idx": 5105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564850993.329, "dur": 2.107,
    "args": {
      "External id": 5107,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2654, "finished": true, "Ev Idx": 5106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564850996.013, "dur": 108.152,
    "args": {
      "External id": 5108,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2655, "finished": true, "Ev Idx": 5107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564851072.955, "dur": 1.342,
    "args": {
      "External id": 5109,"Record function id": 0, "finished": true, "Ev Idx": 5108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564851104.861, "dur": 1.582,
    "args": {
      "External id": 5110,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2656, "finished": true, "Ev Idx": 5109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564851117.242, "dur": 11.697,
    "args": {
      "External id": 5111,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2657, "finished": true, "Ev Idx": 5110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564851137.408, "dur": 112.158,
    "args": {
      "External id": 5112,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2658, "finished": true, "Ev Idx": 5111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564851138.011, "dur": 3.245,
    "args": {
      "External id": 5113,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2658, "finished": true, "Ev Idx": 5112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564851139.342, "dur": 1.249,
    "args": {
      "External id": 5114,"Record function id": 0, "finished": true, "Ev Idx": 5113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851139.958, "dur": 0.494,
    "args": {
      "External id": 5115,"Record function id": 0, "finished": true, "Ev Idx": 5114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564851141.596, "dur": 107.875,
    "args": {
      "External id": 5116,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2659, "finished": true, "Ev Idx": 5115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564851142.188, "dur": 3.050,
    "args": {
      "External id": 5117,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2659, "finished": true, "Ev Idx": 5116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564851142.660, "dur": 2.455,
    "args": {
      "External id": 5118,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2659, "finished": true, "Ev Idx": 5117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564851145.617, "dur": 69.625,
    "args": {
      "External id": 5119,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2660, "finished": true, "Ev Idx": 5118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564851190.633, "dur": 1.371,
    "args": {
      "External id": 5120,"Record function id": 0, "finished": true, "Ev Idx": 5119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564851247.804, "dur": 1.324,
    "args": {
      "External id": 5121,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2661, "finished": true, "Ev Idx": 5120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564851251.799, "dur": 9.334,
    "args": {
      "External id": 5122,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2662, "finished": true, "Ev Idx": 5121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564851268.407, "dur": 42.057,
    "args": {
      "External id": 5123,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2663, "finished": true, "Ev Idx": 5122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564851268.830, "dur": 2.687,
    "args": {
      "External id": 5124,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2663, "finished": true, "Ev Idx": 5123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564851269.812, "dur": 1.101,
    "args": {
      "External id": 5125,"Record function id": 0, "finished": true, "Ev Idx": 5124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851270.325, "dur": 0.459,
    "args": {
      "External id": 5126,"Record function id": 0, "finished": true, "Ev Idx": 5125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564851271.837, "dur": 38.530,
    "args": {
      "External id": 5127,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2664, "finished": true, "Ev Idx": 5126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564851272.255, "dur": 1.878,
    "args": {
      "External id": 5128,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2664, "finished": true, "Ev Idx": 5127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564851272.530, "dur": 1.504,
    "args": {
      "External id": 5129,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2664, "finished": true, "Ev Idx": 5128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564851274.490, "dur": 34.010,
    "args": {
      "External id": 5130,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2665, "finished": true, "Ev Idx": 5129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564851284.368, "dur": 1.108,
    "args": {
      "External id": 5131,"Record function id": 0, "finished": true, "Ev Idx": 5130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564851309.092, "dur": 0.967,
    "args": {
      "External id": 5132,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2666, "finished": true, "Ev Idx": 5131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564851313.305, "dur": 8.664,
    "args": {
      "External id": 5133,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2667, "finished": true, "Ev Idx": 5132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564851343.007, "dur": 0.207,
    "args": {
      "External id": 5134,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2668, "finished": true, "Ev Idx": 5133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564851345.209, "dur": 9.635,
    "args": {
      "External id": 5135,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2668, "finished": true, "Ev Idx": 5134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564851346.523, "dur": 0.492,
    "args": {
      "External id": 5136,"Record function id": 0, "finished": true, "Ev Idx": 5135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564851347.298, "dur": 0.066,
    "args": {
      "External id": 5137,"Record function id": 0, "finished": true, "Ev Idx": 5136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564851357.668, "dur": 13.371,
    "args": {
      "External id": 5138,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2669, "finished": true, "Ev Idx": 5137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564851374.863, "dur": 7.522,
    "args": {
      "External id": 5139,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2670, "finished": true, "Ev Idx": 5138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564851384.061, "dur": 11.595,
    "args": {
      "External id": 5140,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2671, "finished": true, "Ev Idx": 5139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564851397.347, "dur": 8.671,
    "args": {
      "External id": 5141,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2672, "finished": true, "Ev Idx": 5140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564851408.405, "dur": 0.100,
    "args": {
      "External id": 5142,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2673, "finished": true, "Ev Idx": 5141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564851409.467, "dur": 9.214,
    "args": {
      "External id": 5143,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2673, "finished": true, "Ev Idx": 5142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564851436.927, "dur": 37.647,
    "args": {
      "External id": 5144,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2674, "finished": true, "Ev Idx": 5143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564851437.360, "dur": 3.897,
    "args": {
      "External id": 5145,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2674, "finished": true, "Ev Idx": 5144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564851439.284, "dur": 1.213,
    "args": {
      "External id": 5146,"Record function id": 0, "finished": true, "Ev Idx": 5145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851439.793, "dur": 0.530,
    "args": {
      "External id": 5147,"Record function id": 0, "finished": true, "Ev Idx": 5146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564851441.600, "dur": 32.878,
    "args": {
      "External id": 5148,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2675, "finished": true, "Ev Idx": 5147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564851442.049, "dur": 1.455,
    "args": {
      "External id": 5149,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2675, "finished": true, "Ev Idx": 5148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564851442.321, "dur": 1.067,
    "args": {
      "External id": 5150,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2675, "finished": true, "Ev Idx": 5149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564851443.840, "dur": 28.519,
    "args": {
      "External id": 5151,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2676, "finished": true, "Ev Idx": 5150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564851456.787, "dur": 1.059,
    "args": {
      "External id": 5152,"Record function id": 0, "finished": true, "Ev Idx": 5151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564851472.955, "dur": 1.220,
    "args": {
      "External id": 5153,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2677, "finished": true, "Ev Idx": 5152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564851477.664, "dur": 1.296,
    "args": {
      "External id": 5154,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2678, "finished": true, "Ev Idx": 5153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564851480.736, "dur": 2.590,
    "args": {
      "External id": 5155,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2679, "finished": true, "Ev Idx": 5154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851481.847, "dur": 0.580,
    "args": {
      "External id": 5156,"Record function id": 0, "finished": true, "Ev Idx": 5155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564851490.374, "dur": 42.278,
    "args": {
      "External id": 5157,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2680, "finished": true, "Ev Idx": 5156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564851490.726, "dur": 2.238,
    "args": {
      "External id": 5158,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2680, "finished": true, "Ev Idx": 5157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564851491.588, "dur": 0.895,
    "args": {
      "External id": 5159,"Record function id": 0, "finished": true, "Ev Idx": 5158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851492.107, "dur": 0.287,
    "args": {
      "External id": 5160,"Record function id": 0, "finished": true, "Ev Idx": 5159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564851493.270, "dur": 39.287,
    "args": {
      "External id": 5161,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2681, "finished": true, "Ev Idx": 5160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564851493.672, "dur": 1.592,
    "args": {
      "External id": 5162,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2681, "finished": true, "Ev Idx": 5161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564851493.934, "dur": 1.213,
    "args": {
      "External id": 5163,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2681, "finished": true, "Ev Idx": 5162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564851495.564, "dur": 35.118,
    "args": {
      "External id": 5164,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2682, "finished": true, "Ev Idx": 5163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564851508.278, "dur": 1.035,
    "args": {
      "External id": 5165,"Record function id": 0, "finished": true, "Ev Idx": 5164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564851531.203, "dur": 1.079,
    "args": {
      "External id": 5166,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2683, "finished": true, "Ev Idx": 5165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564851534.909, "dur": 1.184,
    "args": {
      "External id": 5167,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2684, "finished": true, "Ev Idx": 5166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564851537.602, "dur": 1.780,
    "args": {
      "External id": 5168,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2685, "finished": true, "Ev Idx": 5167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851538.304, "dur": 0.442,
    "args": {
      "External id": 5169,"Record function id": 0, "finished": true, "Ev Idx": 5168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564851546.521, "dur": 37.268,
    "args": {
      "External id": 5170,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2686, "finished": true, "Ev Idx": 5169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564851546.831, "dur": 2.164,
    "args": {
      "External id": 5171,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2686, "finished": true, "Ev Idx": 5170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564851547.655, "dur": 0.801,
    "args": {
      "External id": 5172,"Record function id": 0, "finished": true, "Ev Idx": 5171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851548.093, "dur": 0.258,
    "args": {
      "External id": 5173,"Record function id": 0, "finished": true, "Ev Idx": 5172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564851549.296, "dur": 34.397,
    "args": {
      "External id": 5174,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2687, "finished": true, "Ev Idx": 5173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564851549.661, "dur": 2.322,
    "args": {
      "External id": 5175,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2687, "finished": true, "Ev Idx": 5174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564851549.928, "dur": 1.960,
    "args": {
      "External id": 5176,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2687, "finished": true, "Ev Idx": 5175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564851552.279, "dur": 29.576,
    "args": {
      "External id": 5177,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2688, "finished": true, "Ev Idx": 5176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564851561.209, "dur": 1.052,
    "args": {
      "External id": 5178,"Record function id": 0, "finished": true, "Ev Idx": 5177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564851582.409, "dur": 1.021,
    "args": {
      "External id": 5179,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2689, "finished": true, "Ev Idx": 5178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564851586.332, "dur": 1.123,
    "args": {
      "External id": 5180,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2690, "finished": true, "Ev Idx": 5179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564851588.795, "dur": 1.599,
    "args": {
      "External id": 5181,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2691, "finished": true, "Ev Idx": 5180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851589.412, "dur": 0.424,
    "args": {
      "External id": 5182,"Record function id": 0, "finished": true, "Ev Idx": 5181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564851593.102, "dur": 2.046,
    "args": {
      "External id": 5183,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2692, "finished": true, "Ev Idx": 5182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851594.293, "dur": 0.319,
    "args": {
      "External id": 5184,"Record function id": 0, "finished": true, "Ev Idx": 5183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564851596.348, "dur": 1.003,
    "args": {
      "External id": 5185,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2692, "finished": true, "Ev Idx": 5184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851596.761, "dur": 0.262,
    "args": {
      "External id": 5186,"Record function id": 0, "finished": true, "Ev Idx": 5185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564851598.239, "dur": 49.871,
    "args": {
      "External id": 5187,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2692, "finished": true, "Ev Idx": 5186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851654.968, "dur": 5.715,
    "args": {
      "External id": 5188,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2693, "finished": true, "Ev Idx": 5187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851658.071, "dur": 0.936,
    "args": {
      "External id": 5189,"Record function id": 0, "finished": true, "Ev Idx": 5188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851663.057, "dur": 3.434,
    "args": {
      "External id": 5190,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2694, "finished": true, "Ev Idx": 5189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851664.980, "dur": 0.536,
    "args": {
      "External id": 5191,"Record function id": 0, "finished": true, "Ev Idx": 5190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564851668.471, "dur": 11.540,
    "args": {
      "External id": 5192,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2695, "finished": true, "Ev Idx": 5191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564851684.688, "dur": 18.576,
    "args": {
      "External id": 5193,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2696, "finished": true, "Ev Idx": 5192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564851706.445, "dur": 9.324,
    "args": {
      "External id": 5194,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2697, "finished": true, "Ev Idx": 5193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564851717.403, "dur": 8.892,
    "args": {
      "External id": 5195,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2698, "finished": true, "Ev Idx": 5194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564851728.249, "dur": 8.793,
    "args": {
      "External id": 5196,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2699, "finished": true, "Ev Idx": 5195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851739.643, "dur": 2.803,
    "args": {
      "External id": 5197,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2700, "finished": true, "Ev Idx": 5196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851741.230, "dur": 0.557,
    "args": {
      "External id": 5198,"Record function id": 0, "finished": true, "Ev Idx": 5197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851744.056, "dur": 1.690,
    "args": {
      "External id": 5199,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2701, "finished": true, "Ev Idx": 5198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851744.890, "dur": 0.323,
    "args": {
      "External id": 5200,"Record function id": 0, "finished": true, "Ev Idx": 5199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564851746.782, "dur": 7.213,
    "args": {
      "External id": 5201,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2702, "finished": true, "Ev Idx": 5200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564851755.996, "dur": 8.831,
    "args": {
      "External id": 5202,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2703, "finished": true, "Ev Idx": 5201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564851767.075, "dur": 6.978,
    "args": {
      "External id": 5203,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2704, "finished": true, "Ev Idx": 5202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564851775.385, "dur": 7.422,
    "args": {
      "External id": 5204,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2705, "finished": true, "Ev Idx": 5203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851807.730, "dur": 2.887,
    "args": {
      "External id": 5205,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2706, "finished": true, "Ev Idx": 5204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851809.401, "dur": 0.475,
    "args": {
      "External id": 5206,"Record function id": 0, "finished": true, "Ev Idx": 5205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851811.353, "dur": 2.333,
    "args": {
      "External id": 5207,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2707, "finished": true, "Ev Idx": 5206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851812.871, "dur": 0.302,
    "args": {
      "External id": 5208,"Record function id": 0, "finished": true, "Ev Idx": 5207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564851814.438, "dur": 2.197,
    "args": {
      "External id": 5209,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2708, "finished": true, "Ev Idx": 5208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851815.392, "dur": 0.606,
    "args": {
      "External id": 5210,"Record function id": 0, "finished": true, "Ev Idx": 5209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851817.443, "dur": 1.788,
    "args": {
      "External id": 5211,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2709, "finished": true, "Ev Idx": 5210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851818.380, "dur": 0.361,
    "args": {
      "External id": 5212,"Record function id": 0, "finished": true, "Ev Idx": 5211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851819.838, "dur": 1.628,
    "args": {
      "External id": 5213,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2710, "finished": true, "Ev Idx": 5212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851820.500, "dur": 0.480,
    "args": {
      "External id": 5214,"Record function id": 0, "finished": true, "Ev Idx": 5213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564851824.231, "dur": 3.440,
    "args": {
      "External id": 5215,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2711, "finished": true, "Ev Idx": 5214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851826.405, "dur": 0.565,
    "args": {
      "External id": 5216,"Record function id": 0, "finished": true, "Ev Idx": 5215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564851830.219, "dur": 21.458,
    "args": {
      "External id": 5217,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2712, "finished": true, "Ev Idx": 5216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564851832.690, "dur": 16.105,
    "args": {
      "External id": 5218,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2712, "finished": true, "Ev Idx": 5217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564851834.757, "dur": 2.770,
    "args": {
      "External id": 5219,"Record function id": 0, "finished": true, "Ev Idx": 5218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564851835.650, "dur": 1.656,
    "args": {
      "External id": 5220,"Record function id": 0, "finished": true, "Ev Idx": 5219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564851838.192, "dur": 10.041,
    "args": {
      "External id": 5221,"Record function id": 0, "finished": true, "Ev Idx": 5220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564851849.350, "dur": 2.070,
    "args": {
      "External id": 5222,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2713, "finished": true, "Ev Idx": 5221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851854.833, "dur": 2.243,
    "args": {
      "External id": 5223,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2714, "finished": true, "Ev Idx": 5222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851856.132, "dur": 0.387,
    "args": {
      "External id": 5224,"Record function id": 0, "finished": true, "Ev Idx": 5223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851857.670, "dur": 1.501,
    "args": {
      "External id": 5225,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2715, "finished": true, "Ev Idx": 5224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851858.374, "dur": 0.361,
    "args": {
      "External id": 5226,"Record function id": 0, "finished": true, "Ev Idx": 5225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564851859.681, "dur": 1.420,
    "args": {
      "External id": 5227,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2716, "finished": true, "Ev Idx": 5226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851860.422, "dur": 0.302,
    "args": {
      "External id": 5228,"Record function id": 0, "finished": true, "Ev Idx": 5227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851861.695, "dur": 3.084,
    "args": {
      "External id": 5229,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2717, "finished": true, "Ev Idx": 5228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851863.877, "dur": 0.461,
    "args": {
      "External id": 5230,"Record function id": 0, "finished": true, "Ev Idx": 5229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851865.374, "dur": 2.199,
    "args": {
      "External id": 5231,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2718, "finished": true, "Ev Idx": 5230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851866.836, "dur": 0.286,
    "args": {
      "External id": 5232,"Record function id": 0, "finished": true, "Ev Idx": 5231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564851869.076, "dur": 1.555,
    "args": {
      "External id": 5233,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2719, "finished": true, "Ev Idx": 5232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851869.896, "dur": 0.323,
    "args": {
      "External id": 5234,"Record function id": 0, "finished": true, "Ev Idx": 5233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564851872.026, "dur": 12.154,
    "args": {
      "External id": 5235,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2720, "finished": true, "Ev Idx": 5234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564851872.915, "dur": 9.812,
    "args": {
      "External id": 5236,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2720, "finished": true, "Ev Idx": 5235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564851873.618, "dur": 2.096,
    "args": {
      "External id": 5237,"Record function id": 0, "finished": true, "Ev Idx": 5236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564851873.966, "dur": 1.640,
    "args": {
      "External id": 5238,"Record function id": 0, "finished": true, "Ev Idx": 5237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564851876.020, "dur": 6.344,
    "args": {
      "External id": 5239,"Record function id": 0, "finished": true, "Ev Idx": 5238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564851883.107, "dur": 0.886,
    "args": {
      "External id": 5240,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2721, "finished": true, "Ev Idx": 5239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564851886.254, "dur": 2.890,
    "args": {
      "External id": 5241,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2722, "finished": true, "Ev Idx": 5240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851888.224, "dur": 0.367,
    "args": {
      "External id": 5242,"Record function id": 0, "finished": true, "Ev Idx": 5241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564851890.572, "dur": 44.911,
    "args": {
      "External id": 5243,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2723, "finished": true, "Ev Idx": 5242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564851891.263, "dur": 2.137,
    "args": {
      "External id": 5244,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2723, "finished": true, "Ev Idx": 5243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851892.683, "dur": 0.315,
    "args": {
      "External id": 5245,"Record function id": 0, "finished": true, "Ev Idx": 5244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564851893.728, "dur": 3.621,
    "args": {
      "External id": 5246,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2724, "finished": true, "Ev Idx": 5245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564851895.366, "dur": 1.774,
    "args": {
      "External id": 5247,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2724, "finished": true, "Ev Idx": 5246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564851897.812, "dur": 1.307,
    "args": {
      "External id": 5248,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2725, "finished": true, "Ev Idx": 5247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851898.461, "dur": 0.279,
    "args": {
      "External id": 5249,"Record function id": 0, "finished": true, "Ev Idx": 5248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564851899.437, "dur": 1.767,
    "args": {
      "External id": 5250,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2726, "finished": true, "Ev Idx": 5249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564851900.313, "dur": 0.804,
    "args": {
      "External id": 5251,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2726, "finished": true, "Ev Idx": 5250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564851901.804, "dur": 31.609,
    "args": {
      "External id": 5252,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2727, "finished": true, "Ev Idx": 5251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564851915.686, "dur": 1.080,
    "args": {
      "External id": 5253,"Record function id": 0, "finished": true, "Ev Idx": 5252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564851933.974, "dur": 1.131,
    "args": {
      "External id": 5254,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2728, "finished": true, "Ev Idx": 5253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564851938.904, "dur": 10.528,
    "args": {
      "External id": 5255,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2729, "finished": true, "Ev Idx": 5254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851952.465, "dur": 1.768,
    "args": {
      "External id": 5256,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2730, "finished": true, "Ev Idx": 5255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851953.188, "dur": 0.450,
    "args": {
      "External id": 5257,"Record function id": 0, "finished": true, "Ev Idx": 5256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851954.910, "dur": 1.258,
    "args": {
      "External id": 5258,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2730, "finished": true, "Ev Idx": 5257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851955.472, "dur": 0.296,
    "args": {
      "External id": 5259,"Record function id": 0, "finished": true, "Ev Idx": 5258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564851956.808, "dur": 1.048,
    "args": {
      "External id": 5260,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2730, "finished": true, "Ev Idx": 5259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564851957.274, "dur": 0.246,
    "args": {
      "External id": 5261,"Record function id": 0, "finished": true, "Ev Idx": 5260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564851959.389, "dur": 7.005,
    "args": {
      "External id": 5262,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2730, "finished": true, "Ev Idx": 5261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564851971.272, "dur": 12.561,
    "args": {
      "External id": 5263,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2731, "finished": true, "Ev Idx": 5262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564851972.165, "dur": 0.389,
    "args": {
      "External id": 5264,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2731, "finished": true, "Ev Idx": 5263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564851973.133, "dur": 10.425,
    "args": {
      "External id": 5265,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2731, "finished": true, "Ev Idx": 5264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564851986.705, "dur": 0.173,
    "args": {
      "External id": 5266,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2732, "finished": true, "Ev Idx": 5265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564851997.585, "dur": 0.478,
    "args": {
      "External id": 5267,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2732, "finished": true, "Ev Idx": 5266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564851999.788, "dur": 37.480,
    "args": {
      "External id": 5268,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2732, "finished": true, "Ev Idx": 5267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564852000.327, "dur": 2.167,
    "args": {
      "External id": 5269,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2732, "finished": true, "Ev Idx": 5268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852001.672, "dur": 0.366,
    "args": {
      "External id": 5270,"Record function id": 0, "finished": true, "Ev Idx": 5269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852002.881, "dur": 1.904,
    "args": {
      "External id": 5271,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2733, "finished": true, "Ev Idx": 5270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852003.203, "dur": 1.457,
    "args": {
      "External id": 5272,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2733, "finished": true, "Ev Idx": 5271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564852005.181, "dur": 2.315,
    "args": {
      "External id": 5273,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2734, "finished": true, "Ev Idx": 5272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852006.743, "dur": 0.414,
    "args": {
      "External id": 5274,"Record function id": 0, "finished": true, "Ev Idx": 5273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852007.789, "dur": 1.851,
    "args": {
      "External id": 5275,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2735, "finished": true, "Ev Idx": 5274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852008.061, "dur": 1.504,
    "args": {
      "External id": 5276,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2735, "finished": true, "Ev Idx": 5275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564852010.036, "dur": 25.345,
    "args": {
      "External id": 5277,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2736, "finished": true, "Ev Idx": 5276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564852019.245, "dur": 1.008,
    "args": {
      "External id": 5278,"Record function id": 0, "finished": true, "Ev Idx": 5277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564852035.960, "dur": 1.011,
    "args": {
      "External id": 5279,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2737, "finished": true, "Ev Idx": 5278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852039.224, "dur": 2.000,
    "args": {
      "External id": 5280,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2738, "finished": true, "Ev Idx": 5279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852040.269, "dur": 0.438,
    "args": {
      "External id": 5281,"Record function id": 0, "finished": true, "Ev Idx": 5280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564852043.603, "dur": 11.835,
    "args": {
      "External id": 5282,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2739, "finished": true, "Ev Idx": 5281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564852044.155, "dur": 11.144,
    "args": {
      "External id": 5283,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2739, "finished": true, "Ev Idx": 5282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564852045.121, "dur": 1.847,
    "args": {
      "External id": 5284,"Record function id": 0, "finished": true, "Ev Idx": 5283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564852045.478, "dur": 1.357,
    "args": {
      "External id": 5285,"Record function id": 0, "finished": true, "Ev Idx": 5284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564852047.253, "dur": 7.651,
    "args": {
      "External id": 5286,"Record function id": 0, "finished": true, "Ev Idx": 5285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852060.641, "dur": 1.610,
    "args": {
      "External id": 5287,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2740, "finished": true, "Ev Idx": 5286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852060.949, "dur": 1.223,
    "args": {
      "External id": 5288,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2740, "finished": true, "Ev Idx": 5287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564852075.665, "dur": 36.335,
    "args": {
      "External id": 5289,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2741, "finished": true, "Ev Idx": 5288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564852076.322, "dur": 2.935,
    "args": {
      "External id": 5290,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2741, "finished": true, "Ev Idx": 5289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852077.700, "dur": 0.916,
    "args": {
      "External id": 5291,"Record function id": 0, "finished": true, "Ev Idx": 5290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852078.174, "dur": 0.322,
    "args": {
      "External id": 5292,"Record function id": 0, "finished": true, "Ev Idx": 5291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564852079.642, "dur": 32.181,
    "args": {
      "External id": 5293,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2742, "finished": true, "Ev Idx": 5292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852080.212, "dur": 2.225,
    "args": {
      "External id": 5294,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2742, "finished": true, "Ev Idx": 5293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852080.664, "dur": 1.661,
    "args": {
      "External id": 5295,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2742, "finished": true, "Ev Idx": 5294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564852082.976, "dur": 26.844,
    "args": {
      "External id": 5296,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2743, "finished": true, "Ev Idx": 5295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564852094.208, "dur": 1.127,
    "args": {
      "External id": 5297,"Record function id": 0, "finished": true, "Ev Idx": 5296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564852110.387, "dur": 1.079,
    "args": {
      "External id": 5298,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2744, "finished": true, "Ev Idx": 5297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564852116.765, "dur": 9.414,
    "args": {
      "External id": 5299,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2745, "finished": true, "Ev Idx": 5298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564852134.900, "dur": 0.262,
    "args": {
      "External id": 5300,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2746, "finished": true, "Ev Idx": 5299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564852137.155, "dur": 11.113,
    "args": {
      "External id": 5301,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2746, "finished": true, "Ev Idx": 5300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564852138.886, "dur": 0.580,
    "args": {
      "External id": 5302,"Record function id": 0, "finished": true, "Ev Idx": 5301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564852139.842, "dur": 0.276,
    "args": {
      "External id": 5303,"Record function id": 0, "finished": true, "Ev Idx": 5302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564852150.806, "dur": 11.405,
    "args": {
      "External id": 5304,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2747, "finished": true, "Ev Idx": 5303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564852165.633, "dur": 9.847,
    "args": {
      "External id": 5305,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2748, "finished": true, "Ev Idx": 5304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564852177.421, "dur": 11.392,
    "args": {
      "External id": 5306,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2749, "finished": true, "Ev Idx": 5305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564852190.682, "dur": 7.536,
    "args": {
      "External id": 5307,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2750, "finished": true, "Ev Idx": 5306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564852200.783, "dur": 0.235,
    "args": {
      "External id": 5308,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2751, "finished": true, "Ev Idx": 5307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564852201.872, "dur": 8.095,
    "args": {
      "External id": 5309,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2751, "finished": true, "Ev Idx": 5308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564852224.742, "dur": 48.094,
    "args": {
      "External id": 5310,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2752, "finished": true, "Ev Idx": 5309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564852225.270, "dur": 3.127,
    "args": {
      "External id": 5311,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2752, "finished": true, "Ev Idx": 5310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852226.496, "dur": 1.333,
    "args": {
      "External id": 5312,"Record function id": 0, "finished": true, "Ev Idx": 5311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852227.144, "dur": 0.572,
    "args": {
      "External id": 5313,"Record function id": 0, "finished": true, "Ev Idx": 5312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564852228.759, "dur": 43.948,
    "args": {
      "External id": 5314,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2753, "finished": true, "Ev Idx": 5313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852229.289, "dur": 3.036,
    "args": {
      "External id": 5315,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2753, "finished": true, "Ev Idx": 5314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852229.652, "dur": 2.565,
    "args": {
      "External id": 5316,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2753, "finished": true, "Ev Idx": 5315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564852232.657, "dur": 38.032,
    "args": {
      "External id": 5317,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2754, "finished": true, "Ev Idx": 5316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564852244.076, "dur": 1.231,
    "args": {
      "External id": 5318,"Record function id": 0, "finished": true, "Ev Idx": 5317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564852271.231, "dur": 1.134,
    "args": {
      "External id": 5319,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2755, "finished": true, "Ev Idx": 5318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564852281.556, "dur": 61.973,
    "args": {
      "External id": 5320,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2756, "finished": true, "Ev Idx": 5319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564852351.944, "dur": 45.379,
    "args": {
      "External id": 5321,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2757, "finished": true, "Ev Idx": 5320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564852352.713, "dur": 3.824,
    "args": {
      "External id": 5322,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2757, "finished": true, "Ev Idx": 5321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852354.433, "dur": 1.544,
    "args": {
      "External id": 5323,"Record function id": 0, "finished": true, "Ev Idx": 5322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852355.157, "dur": 0.667,
    "args": {
      "External id": 5324,"Record function id": 0, "finished": true, "Ev Idx": 5323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564852356.897, "dur": 40.321,
    "args": {
      "External id": 5325,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2758, "finished": true, "Ev Idx": 5324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852357.355, "dur": 1.856,
    "args": {
      "External id": 5326,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2758, "finished": true, "Ev Idx": 5325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852357.798, "dur": 1.289,
    "args": {
      "External id": 5327,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2758, "finished": true, "Ev Idx": 5326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564852359.571, "dur": 35.660,
    "args": {
      "External id": 5328,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2759, "finished": true, "Ev Idx": 5327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564852371.382, "dur": 1.320,
    "args": {
      "External id": 5329,"Record function id": 0, "finished": true, "Ev Idx": 5328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564852395.811, "dur": 1.045,
    "args": {
      "External id": 5330,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2760, "finished": true, "Ev Idx": 5329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564852399.073, "dur": 41.350,
    "args": {
      "External id": 5331,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2761, "finished": true, "Ev Idx": 5330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564852447.654, "dur": 43.043,
    "args": {
      "External id": 5332,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2762, "finished": true, "Ev Idx": 5331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564852448.009, "dur": 3.887,
    "args": {
      "External id": 5333,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2762, "finished": true, "Ev Idx": 5332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852449.316, "dur": 1.398,
    "args": {
      "External id": 5334,"Record function id": 0, "finished": true, "Ev Idx": 5333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852449.858, "dur": 0.688,
    "args": {
      "External id": 5335,"Record function id": 0, "finished": true, "Ev Idx": 5334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564852452.195, "dur": 38.383,
    "args": {
      "External id": 5336,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2763, "finished": true, "Ev Idx": 5335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852452.668, "dur": 2.222,
    "args": {
      "External id": 5337,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2763, "finished": true, "Ev Idx": 5336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852453.065, "dur": 1.720,
    "args": {
      "External id": 5338,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2763, "finished": true, "Ev Idx": 5337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564852455.218, "dur": 33.236,
    "args": {
      "External id": 5339,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2764, "finished": true, "Ev Idx": 5338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564852464.406, "dur": 1.231,
    "args": {
      "External id": 5340,"Record function id": 0, "finished": true, "Ev Idx": 5339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564852489.083, "dur": 1.135,
    "args": {
      "External id": 5341,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2765, "finished": true, "Ev Idx": 5340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564852493.543, "dur": 9.114,
    "args": {
      "External id": 5342,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2766, "finished": true, "Ev Idx": 5341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564852523.385, "dur": 0.397,
    "args": {
      "External id": 5343,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2767, "finished": true, "Ev Idx": 5342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564852526.144, "dur": 11.513,
    "args": {
      "External id": 5344,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2767, "finished": true, "Ev Idx": 5343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564852527.665, "dur": 0.422,
    "args": {
      "External id": 5345,"Record function id": 0, "finished": true, "Ev Idx": 5344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564852528.566, "dur": 0.068,
    "args": {
      "External id": 5346,"Record function id": 0, "finished": true, "Ev Idx": 5345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564852540.449, "dur": 11.625,
    "args": {
      "External id": 5347,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2768, "finished": true, "Ev Idx": 5346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564852555.714, "dur": 8.712,
    "args": {
      "External id": 5348,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2769, "finished": true, "Ev Idx": 5347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564852566.225, "dur": 8.815,
    "args": {
      "External id": 5349,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2770, "finished": true, "Ev Idx": 5348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564852576.885, "dur": 7.017,
    "args": {
      "External id": 5350,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2771, "finished": true, "Ev Idx": 5349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564852586.822, "dur": 0.113,
    "args": {
      "External id": 5351,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2772, "finished": true, "Ev Idx": 5350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564852587.900, "dur": 10.400,
    "args": {
      "External id": 5352,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2772, "finished": true, "Ev Idx": 5351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564852617.764, "dur": 39.378,
    "args": {
      "External id": 5353,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2773, "finished": true, "Ev Idx": 5352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564852618.497, "dur": 4.550,
    "args": {
      "External id": 5354,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2773, "finished": true, "Ev Idx": 5353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852619.948, "dur": 1.935,
    "args": {
      "External id": 5355,"Record function id": 0, "finished": true, "Ev Idx": 5354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852621.064, "dur": 0.653,
    "args": {
      "External id": 5356,"Record function id": 0, "finished": true, "Ev Idx": 5355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564852623.527, "dur": 33.497,
    "args": {
      "External id": 5357,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2774, "finished": true, "Ev Idx": 5356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852624.366, "dur": 2.368,
    "args": {
      "External id": 5358,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2774, "finished": true, "Ev Idx": 5357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852624.860, "dur": 1.719,
    "args": {
      "External id": 5359,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2774, "finished": true, "Ev Idx": 5358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564852627.124, "dur": 27.866,
    "args": {
      "External id": 5360,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2775, "finished": true, "Ev Idx": 5359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564852637.648, "dur": 1.206,
    "args": {
      "External id": 5361,"Record function id": 0, "finished": true, "Ev Idx": 5360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564852655.560, "dur": 1.145,
    "args": {
      "External id": 5362,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2776, "finished": true, "Ev Idx": 5361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852660.332, "dur": 1.489,
    "args": {
      "External id": 5363,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2777, "finished": true, "Ev Idx": 5362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852663.668, "dur": 2.523,
    "args": {
      "External id": 5364,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2778, "finished": true, "Ev Idx": 5363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852664.705, "dur": 0.534,
    "args": {
      "External id": 5365,"Record function id": 0, "finished": true, "Ev Idx": 5364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564852673.240, "dur": 39.686,
    "args": {
      "External id": 5366,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2779, "finished": true, "Ev Idx": 5365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564852673.540, "dur": 2.580,
    "args": {
      "External id": 5367,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2779, "finished": true, "Ev Idx": 5366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852674.703, "dur": 0.944,
    "args": {
      "External id": 5368,"Record function id": 0, "finished": true, "Ev Idx": 5367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852675.232, "dur": 0.310,
    "args": {
      "External id": 5369,"Record function id": 0, "finished": true, "Ev Idx": 5368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564852676.449, "dur": 36.380,
    "args": {
      "External id": 5370,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2780, "finished": true, "Ev Idx": 5369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852676.848, "dur": 1.247,
    "args": {
      "External id": 5371,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2780, "finished": true, "Ev Idx": 5370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852677.100, "dur": 0.884,
    "args": {
      "External id": 5372,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2780, "finished": true, "Ev Idx": 5371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564852678.409, "dur": 32.549,
    "args": {
      "External id": 5373,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2781, "finished": true, "Ev Idx": 5372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564852688.422, "dur": 1.079,
    "args": {
      "External id": 5374,"Record function id": 0, "finished": true, "Ev Idx": 5373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564852711.569, "dur": 0.941,
    "args": {
      "External id": 5375,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2782, "finished": true, "Ev Idx": 5374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852715.118, "dur": 1.117,
    "args": {
      "External id": 5376,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2783, "finished": true, "Ev Idx": 5375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852719.239, "dur": 2.437,
    "args": {
      "External id": 5377,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2784, "finished": true, "Ev Idx": 5376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852720.068, "dur": 0.403,
    "args": {
      "External id": 5378,"Record function id": 0, "finished": true, "Ev Idx": 5377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564852728.193, "dur": 38.357,
    "args": {
      "External id": 5379,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2785, "finished": true, "Ev Idx": 5378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564852728.473, "dur": 2.138,
    "args": {
      "External id": 5380,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2785, "finished": true, "Ev Idx": 5379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852729.376, "dur": 0.839,
    "args": {
      "External id": 5381,"Record function id": 0, "finished": true, "Ev Idx": 5380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852729.752, "dur": 0.339,
    "args": {
      "External id": 5382,"Record function id": 0, "finished": true, "Ev Idx": 5381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564852730.908, "dur": 35.545,
    "args": {
      "External id": 5383,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2786, "finished": true, "Ev Idx": 5382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852731.440, "dur": 1.471,
    "args": {
      "External id": 5384,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2786, "finished": true, "Ev Idx": 5383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852731.997, "dur": 0.809,
    "args": {
      "External id": 5385,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2786, "finished": true, "Ev Idx": 5384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564852733.329, "dur": 31.418,
    "args": {
      "External id": 5386,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2787, "finished": true, "Ev Idx": 5385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564852744.324, "dur": 1.044,
    "args": {
      "External id": 5387,"Record function id": 0, "finished": true, "Ev Idx": 5386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564852765.293, "dur": 0.887,
    "args": {
      "External id": 5388,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2788, "finished": true, "Ev Idx": 5387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564852768.688, "dur": 1.146,
    "args": {
      "External id": 5389,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2789, "finished": true, "Ev Idx": 5388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564852771.208, "dur": 1.793,
    "args": {
      "External id": 5390,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2790, "finished": true, "Ev Idx": 5389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852771.961, "dur": 0.409,
    "args": {
      "External id": 5391,"Record function id": 0, "finished": true, "Ev Idx": 5390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564852775.659, "dur": 2.349,
    "args": {
      "External id": 5392,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2791, "finished": true, "Ev Idx": 5391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852777.120, "dur": 0.308,
    "args": {
      "External id": 5393,"Record function id": 0, "finished": true, "Ev Idx": 5392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564852779.169, "dur": 1.265,
    "args": {
      "External id": 5394,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2791, "finished": true, "Ev Idx": 5393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852779.818, "dur": 0.280,
    "args": {
      "External id": 5395,"Record function id": 0, "finished": true, "Ev Idx": 5394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564852781.261, "dur": 10.369,
    "args": {
      "External id": 5396,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2791, "finished": true, "Ev Idx": 5395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852797.075, "dur": 5.411,
    "args": {
      "External id": 5397,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2792, "finished": true, "Ev Idx": 5396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852800.024, "dur": 0.791,
    "args": {
      "External id": 5398,"Record function id": 0, "finished": true, "Ev Idx": 5397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852804.700, "dur": 2.502,
    "args": {
      "External id": 5399,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2793, "finished": true, "Ev Idx": 5398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852806.232, "dur": 0.346,
    "args": {
      "External id": 5400,"Record function id": 0, "finished": true, "Ev Idx": 5399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564852809.053, "dur": 10.677,
    "args": {
      "External id": 5401,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2794, "finished": true, "Ev Idx": 5400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564852824.060, "dur": 16.632,
    "args": {
      "External id": 5402,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2795, "finished": true, "Ev Idx": 5401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564852843.664, "dur": 7.787,
    "args": {
      "External id": 5403,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2796, "finished": true, "Ev Idx": 5402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564852853.135, "dur": 7.749,
    "args": {
      "External id": 5404,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2797, "finished": true, "Ev Idx": 5403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564852863.068, "dur": 8.647,
    "args": {
      "External id": 5405,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2798, "finished": true, "Ev Idx": 5404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852874.251, "dur": 2.183,
    "args": {
      "External id": 5406,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2799, "finished": true, "Ev Idx": 5405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852875.293, "dur": 0.423,
    "args": {
      "External id": 5407,"Record function id": 0, "finished": true, "Ev Idx": 5406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852877.965, "dur": 1.796,
    "args": {
      "External id": 5408,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2800, "finished": true, "Ev Idx": 5407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852878.791, "dur": 0.439,
    "args": {
      "External id": 5409,"Record function id": 0, "finished": true, "Ev Idx": 5408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564852880.778, "dur": 8.078,
    "args": {
      "External id": 5410,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2801, "finished": true, "Ev Idx": 5409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564852890.696, "dur": 11.192,
    "args": {
      "External id": 5411,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2802, "finished": true, "Ev Idx": 5410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564852904.367, "dur": 6.639,
    "args": {
      "External id": 5412,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2803, "finished": true, "Ev Idx": 5411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564852912.521, "dur": 8.651,
    "args": {
      "External id": 5413,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2804, "finished": true, "Ev Idx": 5412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852944.455, "dur": 2.868,
    "args": {
      "External id": 5414,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2805, "finished": true, "Ev Idx": 5413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852946.207, "dur": 0.444,
    "args": {
      "External id": 5415,"Record function id": 0, "finished": true, "Ev Idx": 5414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852947.943, "dur": 2.299,
    "args": {
      "External id": 5416,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2806, "finished": true, "Ev Idx": 5415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852949.435, "dur": 0.303,
    "args": {
      "External id": 5417,"Record function id": 0, "finished": true, "Ev Idx": 5416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564852951.006, "dur": 2.631,
    "args": {
      "External id": 5418,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2807, "finished": true, "Ev Idx": 5417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852952.270, "dur": 0.711,
    "args": {
      "External id": 5419,"Record function id": 0, "finished": true, "Ev Idx": 5418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852954.432, "dur": 1.753,
    "args": {
      "External id": 5420,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2808, "finished": true, "Ev Idx": 5419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852955.249, "dur": 0.397,
    "args": {
      "External id": 5421,"Record function id": 0, "finished": true, "Ev Idx": 5420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852956.968, "dur": 1.941,
    "args": {
      "External id": 5422,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2809, "finished": true, "Ev Idx": 5421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852958.156, "dur": 0.300,
    "args": {
      "External id": 5423,"Record function id": 0, "finished": true, "Ev Idx": 5422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564852961.709, "dur": 3.736,
    "args": {
      "External id": 5424,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2810, "finished": true, "Ev Idx": 5423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852964.317, "dur": 0.401,
    "args": {
      "External id": 5425,"Record function id": 0, "finished": true, "Ev Idx": 5424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564852968.000, "dur": 19.623,
    "args": {
      "External id": 5426,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2811, "finished": true, "Ev Idx": 5425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564852970.643, "dur": 14.816,
    "args": {
      "External id": 5427,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2811, "finished": true, "Ev Idx": 5426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564852972.423, "dur": 2.703,
    "args": {
      "External id": 5428,"Record function id": 0, "finished": true, "Ev Idx": 5427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564852973.210, "dur": 1.671,
    "args": {
      "External id": 5429,"Record function id": 0, "finished": true, "Ev Idx": 5428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564852975.846, "dur": 9.016,
    "args": {
      "External id": 5430,"Record function id": 0, "finished": true, "Ev Idx": 5429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564852986.060, "dur": 1.290,
    "args": {
      "External id": 5431,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2812, "finished": true, "Ev Idx": 5430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852990.708, "dur": 2.279,
    "args": {
      "External id": 5432,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2813, "finished": true, "Ev Idx": 5431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852992.006, "dur": 0.402,
    "args": {
      "External id": 5433,"Record function id": 0, "finished": true, "Ev Idx": 5432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852993.602, "dur": 1.780,
    "args": {
      "External id": 5434,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2814, "finished": true, "Ev Idx": 5433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852994.646, "dur": 0.305,
    "args": {
      "External id": 5435,"Record function id": 0, "finished": true, "Ev Idx": 5434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564852995.883, "dur": 2.853,
    "args": {
      "External id": 5436,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2815, "finished": true, "Ev Idx": 5435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564852997.905, "dur": 0.337,
    "args": {
      "External id": 5437,"Record function id": 0, "finished": true, "Ev Idx": 5436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564852999.331, "dur": 2.395,
    "args": {
      "External id": 5438,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2816, "finished": true, "Ev Idx": 5437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853000.915, "dur": 0.314,
    "args": {
      "External id": 5439,"Record function id": 0, "finished": true, "Ev Idx": 5438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564853002.347, "dur": 1.469,
    "args": {
      "External id": 5440,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2817, "finished": true, "Ev Idx": 5439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853003.043, "dur": 0.316,
    "args": {
      "External id": 5441,"Record function id": 0, "finished": true, "Ev Idx": 5440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564853005.396, "dur": 1.719,
    "args": {
      "External id": 5442,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2818, "finished": true, "Ev Idx": 5441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853006.333, "dur": 0.349,
    "args": {
      "External id": 5443,"Record function id": 0, "finished": true, "Ev Idx": 5442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853008.531, "dur": 12.735,
    "args": {
      "External id": 5444,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2819, "finished": true, "Ev Idx": 5443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564853009.476, "dur": 9.276,
    "args": {
      "External id": 5445,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2819, "finished": true, "Ev Idx": 5444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564853010.100, "dur": 1.675,
    "args": {
      "External id": 5446,"Record function id": 0, "finished": true, "Ev Idx": 5445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853010.434, "dur": 1.214,
    "args": {
      "External id": 5447,"Record function id": 0, "finished": true, "Ev Idx": 5446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564853012.109, "dur": 6.334,
    "args": {
      "External id": 5448,"Record function id": 0, "finished": true, "Ev Idx": 5447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564853019.138, "dur": 1.955,
    "args": {
      "External id": 5449,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2820, "finished": true, "Ev Idx": 5448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853023.423, "dur": 2.910,
    "args": {
      "External id": 5450,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2821, "finished": true, "Ev Idx": 5449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853025.504, "dur": 0.389,
    "args": {
      "External id": 5451,"Record function id": 0, "finished": true, "Ev Idx": 5450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564853027.821, "dur": 44.297,
    "args": {
      "External id": 5452,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2822, "finished": true, "Ev Idx": 5451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564853028.385, "dur": 1.892,
    "args": {
      "External id": 5453,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2822, "finished": true, "Ev Idx": 5452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853029.516, "dur": 0.326,
    "args": {
      "External id": 5454,"Record function id": 0, "finished": true, "Ev Idx": 5453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853030.608, "dur": 3.772,
    "args": {
      "External id": 5455,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2823, "finished": true, "Ev Idx": 5454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564853032.441, "dur": 1.785,
    "args": {
      "External id": 5456,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2823, "finished": true, "Ev Idx": 5455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564853034.856, "dur": 1.397,
    "args": {
      "External id": 5457,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2824, "finished": true, "Ev Idx": 5456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853035.555, "dur": 0.304,
    "args": {
      "External id": 5458,"Record function id": 0, "finished": true, "Ev Idx": 5457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853036.577, "dur": 1.800,
    "args": {
      "External id": 5459,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2825, "finished": true, "Ev Idx": 5458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564853037.424, "dur": 0.858,
    "args": {
      "External id": 5460,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2825, "finished": true, "Ev Idx": 5459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564853038.967, "dur": 31.078,
    "args": {
      "External id": 5461,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2826, "finished": true, "Ev Idx": 5460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853052.696, "dur": 1.022,
    "args": {
      "External id": 5462,"Record function id": 0, "finished": true, "Ev Idx": 5461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564853070.622, "dur": 1.122,
    "args": {
      "External id": 5463,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2827, "finished": true, "Ev Idx": 5462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564853075.224, "dur": 9.549,
    "args": {
      "External id": 5464,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2828, "finished": true, "Ev Idx": 5463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564853087.774, "dur": 1.888,
    "args": {
      "External id": 5465,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2829, "finished": true, "Ev Idx": 5464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853088.650, "dur": 0.387,
    "args": {
      "External id": 5466,"Record function id": 0, "finished": true, "Ev Idx": 5465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564853090.313, "dur": 1.257,
    "args": {
      "External id": 5467,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2829, "finished": true, "Ev Idx": 5466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853090.890, "dur": 0.311,
    "args": {
      "External id": 5468,"Record function id": 0, "finished": true, "Ev Idx": 5467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564853092.271, "dur": 1.045,
    "args": {
      "External id": 5469,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2829, "finished": true, "Ev Idx": 5468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853092.719, "dur": 0.264,
    "args": {
      "External id": 5470,"Record function id": 0, "finished": true, "Ev Idx": 5469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564853094.856, "dur": 7.316,
    "args": {
      "External id": 5471,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2829, "finished": true, "Ev Idx": 5470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564853107.517, "dur": 13.235,
    "args": {
      "External id": 5472,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2830, "finished": true, "Ev Idx": 5471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564853108.623, "dur": 0.407,
    "args": {
      "External id": 5473,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2830, "finished": true, "Ev Idx": 5472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564853110.006, "dur": 10.452,
    "args": {
      "External id": 5474,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2830, "finished": true, "Ev Idx": 5473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564853123.723, "dur": 0.158,
    "args": {
      "External id": 5475,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2831, "finished": true, "Ev Idx": 5474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564853134.882, "dur": 0.490,
    "args": {
      "External id": 5476,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2831, "finished": true, "Ev Idx": 5475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564853137.203, "dur": 38.831,
    "args": {
      "External id": 5477,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2831, "finished": true, "Ev Idx": 5476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564853138.094, "dur": 1.931,
    "args": {
      "External id": 5478,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2831, "finished": true, "Ev Idx": 5477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853139.121, "dur": 0.388,
    "args": {
      "External id": 5479,"Record function id": 0, "finished": true, "Ev Idx": 5478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853140.661, "dur": 2.830,
    "args": {
      "External id": 5480,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2832, "finished": true, "Ev Idx": 5479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853141.173, "dur": 2.220,
    "args": {
      "External id": 5481,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2832, "finished": true, "Ev Idx": 5480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564853144.002, "dur": 1.804,
    "args": {
      "External id": 5482,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2833, "finished": true, "Ev Idx": 5481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853145.236, "dur": 0.244,
    "args": {
      "External id": 5483,"Record function id": 0, "finished": true, "Ev Idx": 5482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853146.101, "dur": 1.326,
    "args": {
      "External id": 5484,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2834, "finished": true, "Ev Idx": 5483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853146.348, "dur": 0.996,
    "args": {
      "External id": 5485,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2834, "finished": true, "Ev Idx": 5484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564853147.806, "dur": 26.224,
    "args": {
      "External id": 5486,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2835, "finished": true, "Ev Idx": 5485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853158.248, "dur": 1.067,
    "args": {
      "External id": 5487,"Record function id": 0, "finished": true, "Ev Idx": 5486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564853174.680, "dur": 1.018,
    "args": {
      "External id": 5488,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2836, "finished": true, "Ev Idx": 5487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853178.058, "dur": 1.799,
    "args": {
      "External id": 5489,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2837, "finished": true, "Ev Idx": 5488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853179.000, "dur": 0.378,
    "args": {
      "External id": 5490,"Record function id": 0, "finished": true, "Ev Idx": 5489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564853182.379, "dur": 12.032,
    "args": {
      "External id": 5491,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2838, "finished": true, "Ev Idx": 5490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564853182.917, "dur": 11.355,
    "args": {
      "External id": 5492,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2838, "finished": true, "Ev Idx": 5491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564853183.964, "dur": 1.861,
    "args": {
      "External id": 5493,"Record function id": 0, "finished": true, "Ev Idx": 5492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853184.370, "dur": 1.303,
    "args": {
      "External id": 5494,"Record function id": 0, "finished": true, "Ev Idx": 5493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564853186.166, "dur": 7.693,
    "args": {
      "External id": 5495,"Record function id": 0, "finished": true, "Ev Idx": 5494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853199.595, "dur": 1.653,
    "args": {
      "External id": 5496,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2839, "finished": true, "Ev Idx": 5495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853199.889, "dur": 1.253,
    "args": {
      "External id": 5497,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2839, "finished": true, "Ev Idx": 5496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564853213.000, "dur": 38.570,
    "args": {
      "External id": 5498,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2840, "finished": true, "Ev Idx": 5497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564853213.435, "dur": 2.969,
    "args": {
      "External id": 5499,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2840, "finished": true, "Ev Idx": 5498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853214.670, "dur": 1.161,
    "args": {
      "External id": 5500,"Record function id": 0, "finished": true, "Ev Idx": 5499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853215.370, "dur": 0.347,
    "args": {
      "External id": 5501,"Record function id": 0, "finished": true, "Ev Idx": 5500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564853217.283, "dur": 34.145,
    "args": {
      "External id": 5502,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2841, "finished": true, "Ev Idx": 5501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853218.089, "dur": 1.550,
    "args": {
      "External id": 5503,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2841, "finished": true, "Ev Idx": 5502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853218.498, "dur": 1.035,
    "args": {
      "External id": 5504,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2841, "finished": true, "Ev Idx": 5503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564853220.100, "dur": 29.244,
    "args": {
      "External id": 5505,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2842, "finished": true, "Ev Idx": 5504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853232.452, "dur": 1.142,
    "args": {
      "External id": 5506,"Record function id": 0, "finished": true, "Ev Idx": 5505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564853250.104, "dur": 0.978,
    "args": {
      "External id": 5507,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2843, "finished": true, "Ev Idx": 5506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564853256.136, "dur": 65.621,
    "args": {
      "External id": 5508,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2844, "finished": true, "Ev Idx": 5507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564853333.689, "dur": 0.217,
    "args": {
      "External id": 5509,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2845, "finished": true, "Ev Idx": 5508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564853336.142, "dur": 12.586,
    "args": {
      "External id": 5510,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2845, "finished": true, "Ev Idx": 5509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564853338.244, "dur": 0.581,
    "args": {
      "External id": 5511,"Record function id": 0, "finished": true, "Ev Idx": 5510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564853339.281, "dur": 0.102,
    "args": {
      "External id": 5512,"Record function id": 0, "finished": true, "Ev Idx": 5511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564853351.558, "dur": 14.572,
    "args": {
      "External id": 5513,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2846, "finished": true, "Ev Idx": 5512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564853369.981, "dur": 8.452,
    "args": {
      "External id": 5514,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2847, "finished": true, "Ev Idx": 5513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564853380.483, "dur": 10.024,
    "args": {
      "External id": 5515,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2848, "finished": true, "Ev Idx": 5514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564853392.337, "dur": 10.401,
    "args": {
      "External id": 5516,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2849, "finished": true, "Ev Idx": 5515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564853405.201, "dur": 0.229,
    "args": {
      "External id": 5517,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2850, "finished": true, "Ev Idx": 5516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564853406.552, "dur": 9.290,
    "args": {
      "External id": 5518,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2850, "finished": true, "Ev Idx": 5517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564853430.092, "dur": 52.756,
    "args": {
      "External id": 5519,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2851, "finished": true, "Ev Idx": 5518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564853430.934, "dur": 4.205,
    "args": {
      "External id": 5520,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2851, "finished": true, "Ev Idx": 5519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853432.636, "dur": 1.727,
    "args": {
      "External id": 5521,"Record function id": 0, "finished": true, "Ev Idx": 5520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853433.666, "dur": 0.508,
    "args": {
      "External id": 5522,"Record function id": 0, "finished": true, "Ev Idx": 5521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564853435.667, "dur": 47.043,
    "args": {
      "External id": 5523,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2852, "finished": true, "Ev Idx": 5522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853436.246, "dur": 1.753,
    "args": {
      "External id": 5524,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2852, "finished": true, "Ev Idx": 5523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853436.653, "dur": 1.192,
    "args": {
      "External id": 5525,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2852, "finished": true, "Ev Idx": 5524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564853438.381, "dur": 42.220,
    "args": {
      "External id": 5526,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2853, "finished": true, "Ev Idx": 5525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853452.152, "dur": 1.273,
    "args": {
      "External id": 5527,"Record function id": 0, "finished": true, "Ev Idx": 5526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564853481.324, "dur": 1.049,
    "args": {
      "External id": 5528,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2854, "finished": true, "Ev Idx": 5527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564853490.865, "dur": 9.874,
    "args": {
      "External id": 5529,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2855, "finished": true, "Ev Idx": 5528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564853508.129, "dur": 86.985,
    "args": {
      "External id": 5530,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2856, "finished": true, "Ev Idx": 5529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564853508.562, "dur": 3.050,
    "args": {
      "External id": 5531,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2856, "finished": true, "Ev Idx": 5530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853509.509, "dur": 1.550,
    "args": {
      "External id": 5532,"Record function id": 0, "finished": true, "Ev Idx": 5531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853510.505, "dur": 0.427,
    "args": {
      "External id": 5533,"Record function id": 0, "finished": true, "Ev Idx": 5532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564853512.149, "dur": 82.876,
    "args": {
      "External id": 5534,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2857, "finished": true, "Ev Idx": 5533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853513.002, "dur": 3.282,
    "args": {
      "External id": 5535,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2857, "finished": true, "Ev Idx": 5534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853513.433, "dur": 2.738,
    "args": {
      "External id": 5536,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2857, "finished": true, "Ev Idx": 5535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564853516.626, "dur": 76.324,
    "args": {
      "External id": 5537,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2858, "finished": true, "Ev Idx": 5536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853568.132, "dur": 1.192,
    "args": {
      "External id": 5538,"Record function id": 0, "finished": true, "Ev Idx": 5537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564853593.534, "dur": 1.136,
    "args": {
      "External id": 5539,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2859, "finished": true, "Ev Idx": 5538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564853597.100, "dur": 9.712,
    "args": {
      "External id": 5540,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2860, "finished": true, "Ev Idx": 5539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564853613.669, "dur": 52.695,
    "args": {
      "External id": 5541,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2861, "finished": true, "Ev Idx": 5540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564853614.380, "dur": 2.967,
    "args": {
      "External id": 5542,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2861, "finished": true, "Ev Idx": 5541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853615.340, "dur": 1.374,
    "args": {
      "External id": 5543,"Record function id": 0, "finished": true, "Ev Idx": 5542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853615.962, "dur": 0.615,
    "args": {
      "External id": 5544,"Record function id": 0, "finished": true, "Ev Idx": 5543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564853617.656, "dur": 48.588,
    "args": {
      "External id": 5545,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2862, "finished": true, "Ev Idx": 5544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853618.087, "dur": 2.201,
    "args": {
      "External id": 5546,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2862, "finished": true, "Ev Idx": 5545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853618.610, "dur": 1.545,
    "args": {
      "External id": 5547,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2862, "finished": true, "Ev Idx": 5546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564853620.601, "dur": 43.837,
    "args": {
      "External id": 5548,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2863, "finished": true, "Ev Idx": 5547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853632.188, "dur": 1.255,
    "args": {
      "External id": 5549,"Record function id": 0, "finished": true, "Ev Idx": 5548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564853664.964, "dur": 0.969,
    "args": {
      "External id": 5550,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2864, "finished": true, "Ev Idx": 5549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564853668.983, "dur": 8.544,
    "args": {
      "External id": 5551,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2865, "finished": true, "Ev Idx": 5550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564853697.286, "dur": 0.346,
    "args": {
      "External id": 5552,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2866, "finished": true, "Ev Idx": 5551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564853699.521, "dur": 9.168,
    "args": {
      "External id": 5553,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2866, "finished": true, "Ev Idx": 5552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564853700.793, "dur": 0.422,
    "args": {
      "External id": 5554,"Record function id": 0, "finished": true, "Ev Idx": 5553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564853701.569, "dur": 0.072,
    "args": {
      "External id": 5555,"Record function id": 0, "finished": true, "Ev Idx": 5554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564853711.130, "dur": 10.279,
    "args": {
      "External id": 5556,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2867, "finished": true, "Ev Idx": 5555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564853724.984, "dur": 8.630,
    "args": {
      "External id": 5557,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2868, "finished": true, "Ev Idx": 5556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564853735.338, "dur": 8.168,
    "args": {
      "External id": 5558,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2869, "finished": true, "Ev Idx": 5557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564853745.117, "dur": 7.068,
    "args": {
      "External id": 5559,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2870, "finished": true, "Ev Idx": 5558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564853754.733, "dur": 0.273,
    "args": {
      "External id": 5560,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2871, "finished": true, "Ev Idx": 5559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564853756.002, "dur": 42.013,
    "args": {
      "External id": 5561,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2871, "finished": true, "Ev Idx": 5560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564853816.691, "dur": 36.658,
    "args": {
      "External id": 5562,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2872, "finished": true, "Ev Idx": 5561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564853817.683, "dur": 4.038,
    "args": {
      "External id": 5563,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2872, "finished": true, "Ev Idx": 5562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853819.136, "dur": 1.857,
    "args": {
      "External id": 5564,"Record function id": 0, "finished": true, "Ev Idx": 5563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853820.152, "dur": 0.603,
    "args": {
      "External id": 5565,"Record function id": 0, "finished": true, "Ev Idx": 5564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564853822.145, "dur": 31.104,
    "args": {
      "External id": 5566,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2873, "finished": true, "Ev Idx": 5565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853822.749, "dur": 1.942,
    "args": {
      "External id": 5567,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2873, "finished": true, "Ev Idx": 5566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853823.043, "dur": 1.489,
    "args": {
      "External id": 5568,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2873, "finished": true, "Ev Idx": 5567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564853825.021, "dur": 25.968,
    "args": {
      "External id": 5569,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2874, "finished": true, "Ev Idx": 5568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853834.413, "dur": 1.354,
    "args": {
      "External id": 5570,"Record function id": 0, "finished": true, "Ev Idx": 5569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564853851.639, "dur": 1.249,
    "args": {
      "External id": 5571,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2875, "finished": true, "Ev Idx": 5570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853856.620, "dur": 1.345,
    "args": {
      "External id": 5572,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2876, "finished": true, "Ev Idx": 5571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853859.916, "dur": 2.395,
    "args": {
      "External id": 5573,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2877, "finished": true, "Ev Idx": 5572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853861.009, "dur": 0.523,
    "args": {
      "External id": 5574,"Record function id": 0, "finished": true, "Ev Idx": 5573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564853869.273, "dur": 38.867,
    "args": {
      "External id": 5575,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2878, "finished": true, "Ev Idx": 5574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564853869.574, "dur": 2.821,
    "args": {
      "External id": 5576,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2878, "finished": true, "Ev Idx": 5575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853870.550, "dur": 1.370,
    "args": {
      "External id": 5577,"Record function id": 0, "finished": true, "Ev Idx": 5576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853871.245, "dur": 0.562,
    "args": {
      "External id": 5578,"Record function id": 0, "finished": true, "Ev Idx": 5577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564853872.719, "dur": 35.323,
    "args": {
      "External id": 5579,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2879, "finished": true, "Ev Idx": 5578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853873.128, "dur": 1.425,
    "args": {
      "External id": 5580,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2879, "finished": true, "Ev Idx": 5579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853873.400, "dur": 1.058,
    "args": {
      "External id": 5581,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2879, "finished": true, "Ev Idx": 5580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564853874.852, "dur": 31.349,
    "args": {
      "External id": 5582,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2880, "finished": true, "Ev Idx": 5581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853881.922, "dur": 1.298,
    "args": {
      "External id": 5583,"Record function id": 0, "finished": true, "Ev Idx": 5582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564853906.761, "dur": 0.992,
    "args": {
      "External id": 5584,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2881, "finished": true, "Ev Idx": 5583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853910.349, "dur": 1.180,
    "args": {
      "External id": 5585,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2882, "finished": true, "Ev Idx": 5584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853912.952, "dur": 1.975,
    "args": {
      "External id": 5586,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2883, "finished": true, "Ev Idx": 5585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853913.855, "dur": 0.414,
    "args": {
      "External id": 5587,"Record function id": 0, "finished": true, "Ev Idx": 5586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564853921.931, "dur": 34.027,
    "args": {
      "External id": 5588,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2884, "finished": true, "Ev Idx": 5587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564853922.473, "dur": 2.507,
    "args": {
      "External id": 5589,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2884, "finished": true, "Ev Idx": 5588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853923.495, "dur": 1.034,
    "args": {
      "External id": 5590,"Record function id": 0, "finished": true, "Ev Idx": 5589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853924.155, "dur": 0.263,
    "args": {
      "External id": 5591,"Record function id": 0, "finished": true, "Ev Idx": 5590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564853925.301, "dur": 30.578,
    "args": {
      "External id": 5592,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2885, "finished": true, "Ev Idx": 5591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564853925.690, "dur": 1.169,
    "args": {
      "External id": 5593,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2885, "finished": true, "Ev Idx": 5592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853925.955, "dur": 0.811,
    "args": {
      "External id": 5594,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2885, "finished": true, "Ev Idx": 5593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564853927.160, "dur": 27.023,
    "args": {
      "External id": 5595,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2886, "finished": true, "Ev Idx": 5594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564853934.023, "dur": 1.154,
    "args": {
      "External id": 5596,"Record function id": 0, "finished": true, "Ev Idx": 5595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564853954.669, "dur": 0.929,
    "args": {
      "External id": 5597,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2887, "finished": true, "Ev Idx": 5596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564853958.061, "dur": 1.220,
    "args": {
      "External id": 5598,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2888, "finished": true, "Ev Idx": 5597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564853960.740, "dur": 1.605,
    "args": {
      "External id": 5599,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2889, "finished": true, "Ev Idx": 5598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853961.453, "dur": 0.427,
    "args": {
      "External id": 5600,"Record function id": 0, "finished": true, "Ev Idx": 5599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564853965.173, "dur": 2.303,
    "args": {
      "External id": 5601,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2890, "finished": true, "Ev Idx": 5600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853966.457, "dur": 0.454,
    "args": {
      "External id": 5602,"Record function id": 0, "finished": true, "Ev Idx": 5601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564853968.584, "dur": 1.092,
    "args": {
      "External id": 5603,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2890, "finished": true, "Ev Idx": 5602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853969.071, "dur": 0.270,
    "args": {
      "External id": 5604,"Record function id": 0, "finished": true, "Ev Idx": 5603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564853970.555, "dur": 9.570,
    "args": {
      "External id": 5605,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2890, "finished": true, "Ev Idx": 5604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564853985.654, "dur": 4.104,
    "args": {
      "External id": 5606,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2891, "finished": true, "Ev Idx": 5605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853987.995, "dur": 0.485,
    "args": {
      "External id": 5607,"Record function id": 0, "finished": true, "Ev Idx": 5606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564853991.888, "dur": 1.970,
    "args": {
      "External id": 5608,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2892, "finished": true, "Ev Idx": 5607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564853992.738, "dur": 0.465,
    "args": {
      "External id": 5609,"Record function id": 0, "finished": true, "Ev Idx": 5608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564853995.812, "dur": 9.161,
    "args": {
      "External id": 5610,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2893, "finished": true, "Ev Idx": 5609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564854009.176, "dur": 16.427,
    "args": {
      "External id": 5611,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2894, "finished": true, "Ev Idx": 5610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564854028.671, "dur": 9.493,
    "args": {
      "External id": 5612,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2895, "finished": true, "Ev Idx": 5611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564854039.759, "dur": 7.730,
    "args": {
      "External id": 5613,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2896, "finished": true, "Ev Idx": 5612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564854050.085, "dur": 7.785,
    "args": {
      "External id": 5614,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2897, "finished": true, "Ev Idx": 5613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854060.546, "dur": 2.427,
    "args": {
      "External id": 5615,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2898, "finished": true, "Ev Idx": 5614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854061.861, "dur": 0.472,
    "args": {
      "External id": 5616,"Record function id": 0, "finished": true, "Ev Idx": 5615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854064.577, "dur": 1.842,
    "args": {
      "External id": 5617,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2899, "finished": true, "Ev Idx": 5616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854065.403, "dur": 0.486,
    "args": {
      "External id": 5618,"Record function id": 0, "finished": true, "Ev Idx": 5617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564854067.416, "dur": 26.622,
    "args": {
      "External id": 5619,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2900, "finished": true, "Ev Idx": 5618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564854101.061, "dur": 14.806,
    "args": {
      "External id": 5620,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2901, "finished": true, "Ev Idx": 5619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564854119.286, "dur": 9.160,
    "args": {
      "External id": 5621,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2902, "finished": true, "Ev Idx": 5620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564854130.200, "dur": 9.458,
    "args": {
      "External id": 5622,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2903, "finished": true, "Ev Idx": 5621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854164.971, "dur": 3.244,
    "args": {
      "External id": 5623,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2904, "finished": true, "Ev Idx": 5622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854166.649, "dur": 0.580,
    "args": {
      "External id": 5624,"Record function id": 0, "finished": true, "Ev Idx": 5623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854169.155, "dur": 2.376,
    "args": {
      "External id": 5625,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2905, "finished": true, "Ev Idx": 5624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854170.629, "dur": 0.409,
    "args": {
      "External id": 5626,"Record function id": 0, "finished": true, "Ev Idx": 5625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564854172.440, "dur": 2.455,
    "args": {
      "External id": 5627,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2906, "finished": true, "Ev Idx": 5626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854173.552, "dur": 0.693,
    "args": {
      "External id": 5628,"Record function id": 0, "finished": true, "Ev Idx": 5627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854175.523, "dur": 1.802,
    "args": {
      "External id": 5629,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2907, "finished": true, "Ev Idx": 5628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854176.422, "dur": 0.390,
    "args": {
      "External id": 5630,"Record function id": 0, "finished": true, "Ev Idx": 5629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854177.935, "dur": 1.661,
    "args": {
      "External id": 5631,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2908, "finished": true, "Ev Idx": 5630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854178.856, "dur": 0.272,
    "args": {
      "External id": 5632,"Record function id": 0, "finished": true, "Ev Idx": 5631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564854208.460, "dur": 3.710,
    "args": {
      "External id": 5633,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2909, "finished": true, "Ev Idx": 5632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854211.033, "dur": 0.379,
    "args": {
      "External id": 5634,"Record function id": 0, "finished": true, "Ev Idx": 5633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854214.814, "dur": 22.378,
    "args": {
      "External id": 5635,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2910, "finished": true, "Ev Idx": 5634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564854217.608, "dur": 17.339,
    "args": {
      "External id": 5636,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2910, "finished": true, "Ev Idx": 5635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564854219.784, "dur": 3.396,
    "args": {
      "External id": 5637,"Record function id": 0, "finished": true, "Ev Idx": 5636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564854220.777, "dur": 2.066,
    "args": {
      "External id": 5638,"Record function id": 0, "finished": true, "Ev Idx": 5637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564854223.974, "dur": 10.285,
    "args": {
      "External id": 5639,"Record function id": 0, "finished": true, "Ev Idx": 5638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564854235.528, "dur": 1.395,
    "args": {
      "External id": 5640,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2911, "finished": true, "Ev Idx": 5639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854240.492, "dur": 2.405,
    "args": {
      "External id": 5641,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2912, "finished": true, "Ev Idx": 5640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854241.949, "dur": 0.375,
    "args": {
      "External id": 5642,"Record function id": 0, "finished": true, "Ev Idx": 5641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854243.475, "dur": 1.435,
    "args": {
      "External id": 5643,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2913, "finished": true, "Ev Idx": 5642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854244.184, "dur": 0.294,
    "args": {
      "External id": 5644,"Record function id": 0, "finished": true, "Ev Idx": 5643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564854245.453, "dur": 2.153,
    "args": {
      "External id": 5645,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2914, "finished": true, "Ev Idx": 5644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854246.879, "dur": 0.305,
    "args": {
      "External id": 5646,"Record function id": 0, "finished": true, "Ev Idx": 5645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854248.215, "dur": 1.934,
    "args": {
      "External id": 5647,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2915, "finished": true, "Ev Idx": 5646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854249.405, "dur": 0.293,
    "args": {
      "External id": 5648,"Record function id": 0, "finished": true, "Ev Idx": 5647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854250.744, "dur": 1.354,
    "args": {
      "External id": 5649,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2916, "finished": true, "Ev Idx": 5648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854251.370, "dur": 0.289,
    "args": {
      "External id": 5650,"Record function id": 0, "finished": true, "Ev Idx": 5649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564854253.641, "dur": 1.617,
    "args": {
      "External id": 5651,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2917, "finished": true, "Ev Idx": 5650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854254.478, "dur": 0.321,
    "args": {
      "External id": 5652,"Record function id": 0, "finished": true, "Ev Idx": 5651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854256.697, "dur": 11.967,
    "args": {
      "External id": 5653,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2918, "finished": true, "Ev Idx": 5652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564854257.560, "dur": 8.831,
    "args": {
      "External id": 5654,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2918, "finished": true, "Ev Idx": 5653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564854258.226, "dur": 1.657,
    "args": {
      "External id": 5655,"Record function id": 0, "finished": true, "Ev Idx": 5654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564854258.558, "dur": 1.205,
    "args": {
      "External id": 5656,"Record function id": 0, "finished": true, "Ev Idx": 5655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564854260.187, "dur": 5.804,
    "args": {
      "External id": 5657,"Record function id": 0, "finished": true, "Ev Idx": 5656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564854266.807, "dur": 1.707,
    "args": {
      "External id": 5658,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2919, "finished": true, "Ev Idx": 5657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564854271.079, "dur": 2.573,
    "args": {
      "External id": 5659,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2920, "finished": true, "Ev Idx": 5658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854272.790, "dur": 0.360,
    "args": {
      "External id": 5660,"Record function id": 0, "finished": true, "Ev Idx": 5659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564854275.345, "dur": 52.479,
    "args": {
      "External id": 5661,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2921, "finished": true, "Ev Idx": 5660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564854276.130, "dur": 1.499,
    "args": {
      "External id": 5662,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2921, "finished": true, "Ev Idx": 5661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854276.916, "dur": 0.313,
    "args": {
      "External id": 5663,"Record function id": 0, "finished": true, "Ev Idx": 5662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854278.012, "dur": 4.157,
    "args": {
      "External id": 5664,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2922, "finished": true, "Ev Idx": 5663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564854280.015, "dur": 1.996,
    "args": {
      "External id": 5665,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2922, "finished": true, "Ev Idx": 5664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564854282.659, "dur": 1.419,
    "args": {
      "External id": 5666,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2923, "finished": true, "Ev Idx": 5665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854283.399, "dur": 0.303,
    "args": {
      "External id": 5667,"Record function id": 0, "finished": true, "Ev Idx": 5666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854284.378, "dur": 1.766,
    "args": {
      "External id": 5668,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2924, "finished": true, "Ev Idx": 5667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564854285.215, "dur": 0.823,
    "args": {
      "External id": 5669,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2924, "finished": true, "Ev Idx": 5668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564854286.813, "dur": 38.888,
    "args": {
      "External id": 5670,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2925, "finished": true, "Ev Idx": 5669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564854302.953, "dur": 1.222,
    "args": {
      "External id": 5671,"Record function id": 0, "finished": true, "Ev Idx": 5670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564854326.281, "dur": 1.161,
    "args": {
      "External id": 5672,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2926, "finished": true, "Ev Idx": 5671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564854331.767, "dur": 10.305,
    "args": {
      "External id": 5673,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2927, "finished": true, "Ev Idx": 5672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854345.338, "dur": 1.953,
    "args": {
      "External id": 5674,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2928, "finished": true, "Ev Idx": 5673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854346.131, "dur": 0.476,
    "args": {
      "External id": 5675,"Record function id": 0, "finished": true, "Ev Idx": 5674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854347.956, "dur": 1.269,
    "args": {
      "External id": 5676,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2928, "finished": true, "Ev Idx": 5675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854348.526, "dur": 0.306,
    "args": {
      "External id": 5677,"Record function id": 0, "finished": true, "Ev Idx": 5676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564854349.856, "dur": 1.032,
    "args": {
      "External id": 5678,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2928, "finished": true, "Ev Idx": 5677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854350.302, "dur": 0.257,
    "args": {
      "External id": 5679,"Record function id": 0, "finished": true, "Ev Idx": 5678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564854352.620, "dur": 8.210,
    "args": {
      "External id": 5680,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2928, "finished": true, "Ev Idx": 5679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564854366.717, "dur": 15.090,
    "args": {
      "External id": 5681,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2929, "finished": true, "Ev Idx": 5680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564854367.691, "dur": 0.427,
    "args": {
      "External id": 5682,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2929, "finished": true, "Ev Idx": 5681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564854368.931, "dur": 12.620,
    "args": {
      "External id": 5683,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2929, "finished": true, "Ev Idx": 5682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564854385.054, "dur": 0.222,
    "args": {
      "External id": 5684,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2930, "finished": true, "Ev Idx": 5683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564854396.431, "dur": 0.522,
    "args": {
      "External id": 5685,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2930, "finished": true, "Ev Idx": 5684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564854398.744, "dur": 40.639,
    "args": {
      "External id": 5686,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2930, "finished": true, "Ev Idx": 5685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564854399.229, "dur": 2.122,
    "args": {
      "External id": 5687,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2930, "finished": true, "Ev Idx": 5686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854400.438, "dur": 0.388,
    "args": {
      "External id": 5688,"Record function id": 0, "finished": true, "Ev Idx": 5687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854401.753, "dur": 3.669,
    "args": {
      "External id": 5689,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2931, "finished": true, "Ev Idx": 5688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564854402.111, "dur": 3.193,
    "args": {
      "External id": 5690,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2931, "finished": true, "Ev Idx": 5689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564854405.875, "dur": 1.629,
    "args": {
      "External id": 5691,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2932, "finished": true, "Ev Idx": 5690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854406.945, "dur": 0.212,
    "args": {
      "External id": 5692,"Record function id": 0, "finished": true, "Ev Idx": 5691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854407.770, "dur": 1.146,
    "args": {
      "External id": 5693,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2933, "finished": true, "Ev Idx": 5692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564854408.034, "dur": 0.803,
    "args": {
      "External id": 5694,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2933, "finished": true, "Ev Idx": 5693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564854409.327, "dur": 28.077,
    "args": {
      "External id": 5695,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2934, "finished": true, "Ev Idx": 5694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564854419.707, "dur": 1.172,
    "args": {
      "External id": 5696,"Record function id": 0, "finished": true, "Ev Idx": 5695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564854437.983, "dur": 1.126,
    "args": {
      "External id": 5697,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2935, "finished": true, "Ev Idx": 5696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564854441.348, "dur": 1.811,
    "args": {
      "External id": 5698,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2936, "finished": true, "Ev Idx": 5697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854442.277, "dur": 0.347,
    "args": {
      "External id": 5699,"Record function id": 0, "finished": true, "Ev Idx": 5698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564854446.011, "dur": 11.772,
    "args": {
      "External id": 5700,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2937, "finished": true, "Ev Idx": 5699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564854446.538, "dur": 11.091,
    "args": {
      "External id": 5701,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2937, "finished": true, "Ev Idx": 5700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564854447.511, "dur": 1.984,
    "args": {
      "External id": 5702,"Record function id": 0, "finished": true, "Ev Idx": 5701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564854447.891, "dur": 1.446,
    "args": {
      "External id": 5703,"Record function id": 0, "finished": true, "Ev Idx": 5702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564854449.852, "dur": 7.361,
    "args": {
      "External id": 5704,"Record function id": 0, "finished": true, "Ev Idx": 5703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854463.398, "dur": 1.713,
    "args": {
      "External id": 5705,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2938, "finished": true, "Ev Idx": 5704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564854463.715, "dur": 1.293,
    "args": {
      "External id": 5706,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2938, "finished": true, "Ev Idx": 5705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564854481.843, "dur": 35.721,
    "args": {
      "External id": 5707,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2939, "finished": true, "Ev Idx": 5706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564854482.401, "dur": 3.198,
    "args": {
      "External id": 5708,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2939, "finished": true, "Ev Idx": 5707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564854484.108, "dur": 0.957,
    "args": {
      "External id": 5709,"Record function id": 0, "finished": true, "Ev Idx": 5708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854484.549, "dur": 0.398,
    "args": {
      "External id": 5710,"Record function id": 0, "finished": true, "Ev Idx": 5709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564854485.975, "dur": 31.431,
    "args": {
      "External id": 5711,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2940, "finished": true, "Ev Idx": 5710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854486.413, "dur": 1.378,
    "args": {
      "External id": 5712,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2940, "finished": true, "Ev Idx": 5711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564854486.679, "dur": 0.997,
    "args": {
      "External id": 5713,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2940, "finished": true, "Ev Idx": 5712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564854488.274, "dur": 27.098,
    "args": {
      "External id": 5714,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2941, "finished": true, "Ev Idx": 5713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564854500.143, "dur": 1.144,
    "args": {
      "External id": 5715,"Record function id": 0, "finished": true, "Ev Idx": 5714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564854515.942, "dur": 1.066,
    "args": {
      "External id": 5716,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2942, "finished": true, "Ev Idx": 5715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564854522.835, "dur": 8.711,
    "args": {
      "External id": 5717,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2943, "finished": true, "Ev Idx": 5716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564854541.438, "dur": 0.212,
    "args": {
      "External id": 5718,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2944, "finished": true, "Ev Idx": 5717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564854544.040, "dur": 11.996,
    "args": {
      "External id": 5719,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2944, "finished": true, "Ev Idx": 5718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564854545.760, "dur": 0.387,
    "args": {
      "External id": 5720,"Record function id": 0, "finished": true, "Ev Idx": 5719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564854546.575, "dur": 0.060,
    "args": {
      "External id": 5721,"Record function id": 0, "finished": true, "Ev Idx": 5720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564854558.851, "dur": 14.011,
    "args": {
      "External id": 5722,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2945, "finished": true, "Ev Idx": 5721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564854576.866, "dur": 9.038,
    "args": {
      "External id": 5723,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2946, "finished": true, "Ev Idx": 5722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564854588.136, "dur": 9.866,
    "args": {
      "External id": 5724,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2947, "finished": true, "Ev Idx": 5723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564854599.904, "dur": 7.866,
    "args": {
      "External id": 5725,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2948, "finished": true, "Ev Idx": 5724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564854610.463, "dur": 0.100,
    "args": {
      "External id": 5726,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2949, "finished": true, "Ev Idx": 5725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564854611.556, "dur": 8.815,
    "args": {
      "External id": 5727,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2949, "finished": true, "Ev Idx": 5726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564854634.857, "dur": 118.604,
    "args": {
      "External id": 5728,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2950, "finished": true, "Ev Idx": 5727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564854635.259, "dur": 2.936,
    "args": {
      "External id": 5729,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2950, "finished": true, "Ev Idx": 5728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564854636.340, "dur": 1.236,
    "args": {
      "External id": 5730,"Record function id": 0, "finished": true, "Ev Idx": 5729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854636.974, "dur": 0.451,
    "args": {
      "External id": 5731,"Record function id": 0, "finished": true, "Ev Idx": 5730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564854638.652, "dur": 114.609,
    "args": {
      "External id": 5732,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2951, "finished": true, "Ev Idx": 5731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854639.158, "dur": 2.748,
    "args": {
      "External id": 5733,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2951, "finished": true, "Ev Idx": 5732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564854639.454, "dur": 2.337,
    "args": {
      "External id": 5734,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2951, "finished": true, "Ev Idx": 5733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564854642.254, "dur": 108.539,
    "args": {
      "External id": 5735,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2952, "finished": true, "Ev Idx": 5734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564854718.627, "dur": 1.398,
    "args": {
      "External id": 5736,"Record function id": 0, "finished": true, "Ev Idx": 5735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564854751.479, "dur": 1.338,
    "args": {
      "External id": 5737,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2953, "finished": true, "Ev Idx": 5736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564854763.592, "dur": 11.814,
    "args": {
      "External id": 5738,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2954, "finished": true, "Ev Idx": 5737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564854783.037, "dur": 77.324,
    "args": {
      "External id": 5739,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2955, "finished": true, "Ev Idx": 5738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564854783.490, "dur": 3.274,
    "args": {
      "External id": 5740,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2955, "finished": true, "Ev Idx": 5739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564854784.767, "dur": 1.339,
    "args": {
      "External id": 5741,"Record function id": 0, "finished": true, "Ev Idx": 5740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854785.377, "dur": 0.591,
    "args": {
      "External id": 5742,"Record function id": 0, "finished": true, "Ev Idx": 5741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564854787.227, "dur": 73.005,
    "args": {
      "External id": 5743,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2956, "finished": true, "Ev Idx": 5742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854787.859, "dur": 1.731,
    "args": {
      "External id": 5744,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2956, "finished": true, "Ev Idx": 5743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564854788.183, "dur": 1.289,
    "args": {
      "External id": 5745,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2956, "finished": true, "Ev Idx": 5744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564854789.987, "dur": 68.157,
    "args": {
      "External id": 5746,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2957, "finished": true, "Ev Idx": 5745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564854834.647, "dur": 1.199,
    "args": {
      "External id": 5747,"Record function id": 0, "finished": true, "Ev Idx": 5746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564854858.764, "dur": 1.101,
    "args": {
      "External id": 5748,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2958, "finished": true, "Ev Idx": 5747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564854862.527, "dur": 9.107,
    "args": {
      "External id": 5749,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2959, "finished": true, "Ev Idx": 5748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564854878.881, "dur": 40.458,
    "args": {
      "External id": 5750,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2960, "finished": true, "Ev Idx": 5749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564854879.427, "dur": 2.721,
    "args": {
      "External id": 5751,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2960, "finished": true, "Ev Idx": 5750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564854880.460, "dur": 1.101,
    "args": {
      "External id": 5752,"Record function id": 0, "finished": true, "Ev Idx": 5751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564854880.977, "dur": 0.445,
    "args": {
      "External id": 5753,"Record function id": 0, "finished": true, "Ev Idx": 5752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564854882.436, "dur": 36.775,
    "args": {
      "External id": 5754,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2961, "finished": true, "Ev Idx": 5753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564854882.832, "dur": 1.784,
    "args": {
      "External id": 5755,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2961, "finished": true, "Ev Idx": 5754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564854883.254, "dur": 1.257,
    "args": {
      "External id": 5756,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2961, "finished": true, "Ev Idx": 5755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564854885.040, "dur": 32.349,
    "args": {
      "External id": 5757,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2962, "finished": true, "Ev Idx": 5756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564854895.198, "dur": 1.281,
    "args": {
      "External id": 5758,"Record function id": 0, "finished": true, "Ev Idx": 5757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564854917.989, "dur": 0.908,
    "args": {
      "External id": 5759,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2963, "finished": true, "Ev Idx": 5758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564854922.110, "dur": 8.766,
    "args": {
      "External id": 5760,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2964, "finished": true, "Ev Idx": 5759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564854952.216, "dur": 0.304,
    "args": {
      "External id": 5761,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2965, "finished": true, "Ev Idx": 5760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564854954.596, "dur": 10.387,
    "args": {
      "External id": 5762,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2965, "finished": true, "Ev Idx": 5761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564854955.964, "dur": 0.411,
    "args": {
      "External id": 5763,"Record function id": 0, "finished": true, "Ev Idx": 5762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564854956.681, "dur": 0.090,
    "args": {
      "External id": 5764,"Record function id": 0, "finished": true, "Ev Idx": 5763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564854967.668, "dur": 12.495,
    "args": {
      "External id": 5765,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2966, "finished": true, "Ev Idx": 5764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564854984.080, "dur": 10.143,
    "args": {
      "External id": 5766,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2967, "finished": true, "Ev Idx": 5765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564854996.206, "dur": 8.886,
    "args": {
      "External id": 5767,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2968, "finished": true, "Ev Idx": 5766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564855006.822, "dur": 10.456,
    "args": {
      "External id": 5768,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2969, "finished": true, "Ev Idx": 5767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564855019.772, "dur": 0.093,
    "args": {
      "External id": 5769,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2970, "finished": true, "Ev Idx": 5768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564855020.846, "dur": 8.016,
    "args": {
      "External id": 5770,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2970, "finished": true, "Ev Idx": 5769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564855048.098, "dur": 36.336,
    "args": {
      "External id": 5771,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2971, "finished": true, "Ev Idx": 5770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564855048.580, "dur": 3.746,
    "args": {
      "External id": 5772,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2971, "finished": true, "Ev Idx": 5771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855050.249, "dur": 1.327,
    "args": {
      "External id": 5773,"Record function id": 0, "finished": true, "Ev Idx": 5772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855050.913, "dur": 0.523,
    "args": {
      "External id": 5774,"Record function id": 0, "finished": true, "Ev Idx": 5773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564855052.790, "dur": 31.516,
    "args": {
      "External id": 5775,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2972, "finished": true, "Ev Idx": 5774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855053.419, "dur": 1.698,
    "args": {
      "External id": 5776,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2972, "finished": true, "Ev Idx": 5775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855053.858, "dur": 1.160,
    "args": {
      "External id": 5777,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2972, "finished": true, "Ev Idx": 5776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564855055.466, "dur": 26.658,
    "args": {
      "External id": 5778,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2973, "finished": true, "Ev Idx": 5777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855065.902, "dur": 1.210,
    "args": {
      "External id": 5779,"Record function id": 0, "finished": true, "Ev Idx": 5778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564855082.822, "dur": 1.185,
    "args": {
      "External id": 5780,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2974, "finished": true, "Ev Idx": 5779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855087.904, "dur": 1.474,
    "args": {
      "External id": 5781,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2975, "finished": true, "Ev Idx": 5780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855091.553, "dur": 2.564,
    "args": {
      "External id": 5782,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2976, "finished": true, "Ev Idx": 5781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855092.701, "dur": 0.599,
    "args": {
      "External id": 5783,"Record function id": 0, "finished": true, "Ev Idx": 5782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564855100.712, "dur": 40.134,
    "args": {
      "External id": 5784,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2977, "finished": true, "Ev Idx": 5783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564855101.171, "dur": 2.372,
    "args": {
      "External id": 5785,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2977, "finished": true, "Ev Idx": 5784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855102.099, "dur": 0.985,
    "args": {
      "External id": 5786,"Record function id": 0, "finished": true, "Ev Idx": 5785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855102.682, "dur": 0.285,
    "args": {
      "External id": 5787,"Record function id": 0, "finished": true, "Ev Idx": 5786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564855103.975, "dur": 36.783,
    "args": {
      "External id": 5788,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2978, "finished": true, "Ev Idx": 5787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855104.672, "dur": 1.298,
    "args": {
      "External id": 5789,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2978, "finished": true, "Ev Idx": 5788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855104.937, "dur": 0.930,
    "args": {
      "External id": 5790,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2978, "finished": true, "Ev Idx": 5789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564855106.249, "dur": 32.559,
    "args": {
      "External id": 5791,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2979, "finished": true, "Ev Idx": 5790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855116.140, "dur": 1.115,
    "args": {
      "External id": 5792,"Record function id": 0, "finished": true, "Ev Idx": 5791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564855139.362, "dur": 1.129,
    "args": {
      "External id": 5793,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2980, "finished": true, "Ev Idx": 5792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855143.404, "dur": 1.104,
    "args": {
      "External id": 5794,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2981, "finished": true, "Ev Idx": 5793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855146.053, "dur": 1.960,
    "args": {
      "External id": 5795,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2982, "finished": true, "Ev Idx": 5794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855146.936, "dur": 0.439,
    "args": {
      "External id": 5796,"Record function id": 0, "finished": true, "Ev Idx": 5795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564855154.757, "dur": 39.237,
    "args": {
      "External id": 5797,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2983, "finished": true, "Ev Idx": 5796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564855155.011, "dur": 2.245,
    "args": {
      "External id": 5798,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2983, "finished": true, "Ev Idx": 5797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855155.850, "dur": 0.964,
    "args": {
      "External id": 5799,"Record function id": 0, "finished": true, "Ev Idx": 5798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855156.434, "dur": 0.256,
    "args": {
      "External id": 5800,"Record function id": 0, "finished": true, "Ev Idx": 5799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564855157.574, "dur": 36.311,
    "args": {
      "External id": 5801,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2984, "finished": true, "Ev Idx": 5800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855157.988, "dur": 1.173,
    "args": {
      "External id": 5802,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2984, "finished": true, "Ev Idx": 5801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855158.255, "dur": 0.806,
    "args": {
      "External id": 5803,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2984, "finished": true, "Ev Idx": 5802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564855159.450, "dur": 32.693,
    "args": {
      "External id": 5804,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2985, "finished": true, "Ev Idx": 5803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855171.477, "dur": 1.045,
    "args": {
      "External id": 5805,"Record function id": 0, "finished": true, "Ev Idx": 5804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564855192.716, "dur": 0.917,
    "args": {
      "External id": 5806,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2986, "finished": true, "Ev Idx": 5805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855196.412, "dur": 1.154,
    "args": {
      "External id": 5807,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2987, "finished": true, "Ev Idx": 5806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855199.082, "dur": 1.618,
    "args": {
      "External id": 5808,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2988, "finished": true, "Ev Idx": 5807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855199.786, "dur": 0.373,
    "args": {
      "External id": 5809,"Record function id": 0, "finished": true, "Ev Idx": 5808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564855203.369, "dur": 2.194,
    "args": {
      "External id": 5810,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2989, "finished": true, "Ev Idx": 5809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855204.726, "dur": 0.329,
    "args": {
      "External id": 5811,"Record function id": 0, "finished": true, "Ev Idx": 5810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564855206.753, "dur": 1.155,
    "args": {
      "External id": 5812,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2989, "finished": true, "Ev Idx": 5811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855207.225, "dur": 0.334,
    "args": {
      "External id": 5813,"Record function id": 0, "finished": true, "Ev Idx": 5812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564855208.712, "dur": 9.422,
    "args": {
      "External id": 5814,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2989, "finished": true, "Ev Idx": 5813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855223.831, "dur": 4.827,
    "args": {
      "External id": 5815,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2990, "finished": true, "Ev Idx": 5814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855226.061, "dur": 0.825,
    "args": {
      "External id": 5816,"Record function id": 0, "finished": true, "Ev Idx": 5815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855231.063, "dur": 2.366,
    "args": {
      "External id": 5817,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2991, "finished": true, "Ev Idx": 5816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855232.500, "dur": 0.358,
    "args": {
      "External id": 5818,"Record function id": 0, "finished": true, "Ev Idx": 5817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564855235.561, "dur": 9.365,
    "args": {
      "External id": 5819,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2992, "finished": true, "Ev Idx": 5818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564855249.193, "dur": 57.180,
    "args": {
      "External id": 5820,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2993, "finished": true, "Ev Idx": 5819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564855309.918, "dur": 8.562,
    "args": {
      "External id": 5821,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2994, "finished": true, "Ev Idx": 5820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564855320.293, "dur": 8.653,
    "args": {
      "External id": 5822,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2995, "finished": true, "Ev Idx": 5821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564855331.402, "dur": 7.889,
    "args": {
      "External id": 5823,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2996, "finished": true, "Ev Idx": 5822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855342.433, "dur": 2.663,
    "args": {
      "External id": 5824,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2997, "finished": true, "Ev Idx": 5823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855343.753, "dur": 0.509,
    "args": {
      "External id": 5825,"Record function id": 0, "finished": true, "Ev Idx": 5824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855346.710, "dur": 2.772,
    "args": {
      "External id": 5826,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2998, "finished": true, "Ev Idx": 5825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855348.579, "dur": 0.375,
    "args": {
      "External id": 5827,"Record function id": 0, "finished": true, "Ev Idx": 5826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564855350.680, "dur": 7.680,
    "args": {
      "External id": 5828,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 2999, "finished": true, "Ev Idx": 5827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564855360.712, "dur": 9.863,
    "args": {
      "External id": 5829,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3000, "finished": true, "Ev Idx": 5828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564855372.782, "dur": 8.533,
    "args": {
      "External id": 5830,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3001, "finished": true, "Ev Idx": 5829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564855382.840, "dur": 6.228,
    "args": {
      "External id": 5831,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3002, "finished": true, "Ev Idx": 5830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855412.622, "dur": 2.531,
    "args": {
      "External id": 5832,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3003, "finished": true, "Ev Idx": 5831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855413.901, "dur": 0.457,
    "args": {
      "External id": 5833,"Record function id": 0, "finished": true, "Ev Idx": 5832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855415.848, "dur": 1.612,
    "args": {
      "External id": 5834,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3004, "finished": true, "Ev Idx": 5833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855416.625, "dur": 0.319,
    "args": {
      "External id": 5835,"Record function id": 0, "finished": true, "Ev Idx": 5834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564855418.189, "dur": 3.446,
    "args": {
      "External id": 5836,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3005, "finished": true, "Ev Idx": 5835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855420.392, "dur": 0.606,
    "args": {
      "External id": 5837,"Record function id": 0, "finished": true, "Ev Idx": 5836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855422.301, "dur": 2.686,
    "args": {
      "External id": 5838,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3006, "finished": true, "Ev Idx": 5837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855423.974, "dur": 0.464,
    "args": {
      "External id": 5839,"Record function id": 0, "finished": true, "Ev Idx": 5838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855425.655, "dur": 1.602,
    "args": {
      "External id": 5840,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3007, "finished": true, "Ev Idx": 5839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855426.482, "dur": 0.316,
    "args": {
      "External id": 5841,"Record function id": 0, "finished": true, "Ev Idx": 5840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564855429.957, "dur": 2.995,
    "args": {
      "External id": 5842,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3008, "finished": true, "Ev Idx": 5841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855431.671, "dur": 0.598,
    "args": {
      "External id": 5843,"Record function id": 0, "finished": true, "Ev Idx": 5842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855435.554, "dur": 19.974,
    "args": {
      "External id": 5844,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3009, "finished": true, "Ev Idx": 5843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564855438.071, "dur": 14.520,
    "args": {
      "External id": 5845,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3009, "finished": true, "Ev Idx": 5844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564855439.685, "dur": 2.720,
    "args": {
      "External id": 5846,"Record function id": 0, "finished": true, "Ev Idx": 5845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855440.534, "dur": 1.642,
    "args": {
      "External id": 5847,"Record function id": 0, "finished": true, "Ev Idx": 5846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564855443.111, "dur": 8.883,
    "args": {
      "External id": 5848,"Record function id": 0, "finished": true, "Ev Idx": 5847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564855453.204, "dur": 2.060,
    "args": {
      "External id": 5849,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3010, "finished": true, "Ev Idx": 5848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855459.026, "dur": 2.272,
    "args": {
      "External id": 5850,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3011, "finished": true, "Ev Idx": 5849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855460.348, "dur": 0.383,
    "args": {
      "External id": 5851,"Record function id": 0, "finished": true, "Ev Idx": 5850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855461.869, "dur": 1.739,
    "args": {
      "External id": 5852,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3012, "finished": true, "Ev Idx": 5851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855462.782, "dur": 0.379,
    "args": {
      "External id": 5853,"Record function id": 0, "finished": true, "Ev Idx": 5852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564855464.143, "dur": 1.573,
    "args": {
      "External id": 5854,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3013, "finished": true, "Ev Idx": 5853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855464.887, "dur": 0.418,
    "args": {
      "External id": 5855,"Record function id": 0, "finished": true, "Ev Idx": 5854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855466.271, "dur": 1.479,
    "args": {
      "External id": 5856,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3014, "finished": true, "Ev Idx": 5855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855466.988, "dur": 0.297,
    "args": {
      "External id": 5857,"Record function id": 0, "finished": true, "Ev Idx": 5856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855468.335, "dur": 1.416,
    "args": {
      "External id": 5858,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3015, "finished": true, "Ev Idx": 5857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855468.995, "dur": 0.300,
    "args": {
      "External id": 5859,"Record function id": 0, "finished": true, "Ev Idx": 5858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564855471.279, "dur": 2.210,
    "args": {
      "External id": 5860,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3016, "finished": true, "Ev Idx": 5859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855472.738, "dur": 0.314,
    "args": {
      "External id": 5861,"Record function id": 0, "finished": true, "Ev Idx": 5860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855475.088, "dur": 13.138,
    "args": {
      "External id": 5862,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3017, "finished": true, "Ev Idx": 5861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564855476.199, "dur": 10.453,
    "args": {
      "External id": 5863,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3017, "finished": true, "Ev Idx": 5862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564855477.328, "dur": 2.021,
    "args": {
      "External id": 5864,"Record function id": 0, "finished": true, "Ev Idx": 5863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855477.938, "dur": 1.300,
    "args": {
      "External id": 5865,"Record function id": 0, "finished": true, "Ev Idx": 5864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564855479.660, "dur": 6.613,
    "args": {
      "External id": 5866,"Record function id": 0, "finished": true, "Ev Idx": 5865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564855487.041, "dur": 0.980,
    "args": {
      "External id": 5867,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3018, "finished": true, "Ev Idx": 5866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855490.412, "dur": 1.812,
    "args": {
      "External id": 5868,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3019, "finished": true, "Ev Idx": 5867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855491.321, "dur": 0.368,
    "args": {
      "External id": 5869,"Record function id": 0, "finished": true, "Ev Idx": 5868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564855493.780, "dur": 47.983,
    "args": {
      "External id": 5870,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3020, "finished": true, "Ev Idx": 5869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564855494.432, "dur": 1.495,
    "args": {
      "External id": 5871,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3020, "finished": true, "Ev Idx": 5870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855495.195, "dur": 0.318,
    "args": {
      "External id": 5872,"Record function id": 0, "finished": true, "Ev Idx": 5871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855496.327, "dur": 4.276,
    "args": {
      "External id": 5873,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3021, "finished": true, "Ev Idx": 5872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564855498.212, "dur": 2.260,
    "args": {
      "External id": 5874,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3021, "finished": true, "Ev Idx": 5873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564855501.122, "dur": 2.526,
    "args": {
      "External id": 5875,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3022, "finished": true, "Ev Idx": 5874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855502.979, "dur": 0.279,
    "args": {
      "External id": 5876,"Record function id": 0, "finished": true, "Ev Idx": 5875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855504.002, "dur": 2.651,
    "args": {
      "External id": 5877,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3023, "finished": true, "Ev Idx": 5876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564855504.901, "dur": 1.629,
    "args": {
      "External id": 5878,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3023, "finished": true, "Ev Idx": 5877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564855507.244, "dur": 32.551,
    "args": {
      "External id": 5879,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3024, "finished": true, "Ev Idx": 5878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855520.162, "dur": 1.008,
    "args": {
      "External id": 5880,"Record function id": 0, "finished": true, "Ev Idx": 5879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564855540.347, "dur": 1.122,
    "args": {
      "External id": 5881,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3025, "finished": true, "Ev Idx": 5880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564855545.153, "dur": 9.907,
    "args": {
      "External id": 5882,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3026, "finished": true, "Ev Idx": 5881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855557.994, "dur": 1.781,
    "args": {
      "External id": 5883,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3027, "finished": true, "Ev Idx": 5882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855558.770, "dur": 0.383,
    "args": {
      "External id": 5884,"Record function id": 0, "finished": true, "Ev Idx": 5883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855560.493, "dur": 1.330,
    "args": {
      "External id": 5885,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3027, "finished": true, "Ev Idx": 5884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855561.161, "dur": 0.310,
    "args": {
      "External id": 5886,"Record function id": 0, "finished": true, "Ev Idx": 5885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564855562.593, "dur": 1.076,
    "args": {
      "External id": 5887,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3027, "finished": true, "Ev Idx": 5886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855563.112, "dur": 0.245,
    "args": {
      "External id": 5888,"Record function id": 0, "finished": true, "Ev Idx": 5887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564855565.303, "dur": 7.539,
    "args": {
      "External id": 5889,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3027, "finished": true, "Ev Idx": 5888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564855578.234, "dur": 12.881,
    "args": {
      "External id": 5890,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3028, "finished": true, "Ev Idx": 5889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564855579.087, "dur": 0.413,
    "args": {
      "External id": 5891,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3028, "finished": true, "Ev Idx": 5890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564855580.182, "dur": 10.670,
    "args": {
      "External id": 5892,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3028, "finished": true, "Ev Idx": 5891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564855593.892, "dur": 0.181,
    "args": {
      "External id": 5893,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3029, "finished": true, "Ev Idx": 5892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564855604.532, "dur": 0.578,
    "args": {
      "External id": 5894,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3029, "finished": true, "Ev Idx": 5893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564855606.918, "dur": 43.176,
    "args": {
      "External id": 5895,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3029, "finished": true, "Ev Idx": 5894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564855607.442, "dur": 2.963,
    "args": {
      "External id": 5896,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3029, "finished": true, "Ev Idx": 5895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855609.544, "dur": 0.379,
    "args": {
      "External id": 5897,"Record function id": 0, "finished": true, "Ev Idx": 5896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855611.066, "dur": 3.143,
    "args": {
      "External id": 5898,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3030, "finished": true, "Ev Idx": 5897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855611.431, "dur": 2.670,
    "args": {
      "External id": 5899,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3030, "finished": true, "Ev Idx": 5898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564855614.641, "dur": 1.606,
    "args": {
      "External id": 5900,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3031, "finished": true, "Ev Idx": 5899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855615.655, "dur": 0.266,
    "args": {
      "External id": 5901,"Record function id": 0, "finished": true, "Ev Idx": 5900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855616.523, "dur": 2.030,
    "args": {
      "External id": 5902,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3032, "finished": true, "Ev Idx": 5901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855616.834, "dur": 1.622,
    "args": {
      "External id": 5903,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3032, "finished": true, "Ev Idx": 5902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564855619.023, "dur": 29.066,
    "args": {
      "External id": 5904,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3033, "finished": true, "Ev Idx": 5903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855631.751, "dur": 1.068,
    "args": {
      "External id": 5905,"Record function id": 0, "finished": true, "Ev Idx": 5904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564855648.619, "dur": 1.148,
    "args": {
      "External id": 5906,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3034, "finished": true, "Ev Idx": 5905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855651.950, "dur": 1.782,
    "args": {
      "External id": 5907,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3035, "finished": true, "Ev Idx": 5906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855652.829, "dur": 0.385,
    "args": {
      "External id": 5908,"Record function id": 0, "finished": true, "Ev Idx": 5907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564855656.176, "dur": 12.181,
    "args": {
      "External id": 5909,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3036, "finished": true, "Ev Idx": 5908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564855656.739, "dur": 11.477,
    "args": {
      "External id": 5910,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3036, "finished": true, "Ev Idx": 5909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564855657.822, "dur": 1.804,
    "args": {
      "External id": 5911,"Record function id": 0, "finished": true, "Ev Idx": 5910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855658.217, "dur": 1.206,
    "args": {
      "External id": 5912,"Record function id": 0, "finished": true, "Ev Idx": 5911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564855659.920, "dur": 7.871,
    "args": {
      "External id": 5913,"Record function id": 0, "finished": true, "Ev Idx": 5912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855673.670, "dur": 1.673,
    "args": {
      "External id": 5914,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3037, "finished": true, "Ev Idx": 5913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855674.033, "dur": 1.198,
    "args": {
      "External id": 5915,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3037, "finished": true, "Ev Idx": 5914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564855688.838, "dur": 36.954,
    "args": {
      "External id": 5916,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3038, "finished": true, "Ev Idx": 5915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564855689.451, "dur": 2.907,
    "args": {
      "External id": 5917,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3038, "finished": true, "Ev Idx": 5916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855690.870, "dur": 0.952,
    "args": {
      "External id": 5918,"Record function id": 0, "finished": true, "Ev Idx": 5917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855691.282, "dur": 0.424,
    "args": {
      "External id": 5919,"Record function id": 0, "finished": true, "Ev Idx": 5918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564855692.712, "dur": 32.905,
    "args": {
      "External id": 5920,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3039, "finished": true, "Ev Idx": 5919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855693.166, "dur": 3.100,
    "args": {
      "External id": 5921,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3039, "finished": true, "Ev Idx": 5920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855693.461, "dur": 2.716,
    "args": {
      "External id": 5922,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3039, "finished": true, "Ev Idx": 5921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564855696.685, "dur": 26.857,
    "args": {
      "External id": 5923,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3040, "finished": true, "Ev Idx": 5922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855706.927, "dur": 1.037,
    "args": {
      "External id": 5924,"Record function id": 0, "finished": true, "Ev Idx": 5923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564855724.245, "dur": 1.004,
    "args": {
      "External id": 5925,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3041, "finished": true, "Ev Idx": 5924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564855730.643, "dur": 10.009,
    "args": {
      "External id": 5926,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3042, "finished": true, "Ev Idx": 5925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564855749.145, "dur": 0.237,
    "args": {
      "External id": 5927,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3043, "finished": true, "Ev Idx": 5926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564855751.583, "dur": 11.342,
    "args": {
      "External id": 5928,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3043, "finished": true, "Ev Idx": 5927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564855753.581, "dur": 0.341,
    "args": {
      "External id": 5929,"Record function id": 0, "finished": true, "Ev Idx": 5928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564855754.470, "dur": 0.097,
    "args": {
      "External id": 5930,"Record function id": 0, "finished": true, "Ev Idx": 5929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564855765.539, "dur": 12.948,
    "args": {
      "External id": 5931,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3044, "finished": true, "Ev Idx": 5930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564855781.982, "dur": 7.989,
    "args": {
      "External id": 5932,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3045, "finished": true, "Ev Idx": 5931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564855792.026, "dur": 10.251,
    "args": {
      "External id": 5933,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3046, "finished": true, "Ev Idx": 5932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564855804.083, "dur": 10.428,
    "args": {
      "External id": 5934,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3047, "finished": true, "Ev Idx": 5933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564855817.024, "dur": 0.207,
    "args": {
      "External id": 5935,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3048, "finished": true, "Ev Idx": 5934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564855818.304, "dur": 8.873,
    "args": {
      "External id": 5936,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3048, "finished": true, "Ev Idx": 5935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564855840.944, "dur": 49.547,
    "args": {
      "External id": 5937,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3049, "finished": true, "Ev Idx": 5936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564855841.347, "dur": 2.998,
    "args": {
      "External id": 5938,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3049, "finished": true, "Ev Idx": 5937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855842.538, "dur": 1.167,
    "args": {
      "External id": 5939,"Record function id": 0, "finished": true, "Ev Idx": 5938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855843.173, "dur": 0.403,
    "args": {
      "External id": 5940,"Record function id": 0, "finished": true, "Ev Idx": 5939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564855844.691, "dur": 45.688,
    "args": {
      "External id": 5941,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3050, "finished": true, "Ev Idx": 5940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855845.096, "dur": 1.655,
    "args": {
      "External id": 5942,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3050, "finished": true, "Ev Idx": 5941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855845.589, "dur": 1.061,
    "args": {
      "External id": 5943,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3050, "finished": true, "Ev Idx": 5942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564855847.066, "dur": 41.259,
    "args": {
      "External id": 5944,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3051, "finished": true, "Ev Idx": 5943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855861.735, "dur": 1.195,
    "args": {
      "External id": 5945,"Record function id": 0, "finished": true, "Ev Idx": 5944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564855888.978, "dur": 1.093,
    "args": {
      "External id": 5946,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3052, "finished": true, "Ev Idx": 5945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564855898.942, "dur": 61.212,
    "args": {
      "External id": 5947,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3053, "finished": true, "Ev Idx": 5946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564855969.954, "dur": 47.565,
    "args": {
      "External id": 5948,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3054, "finished": true, "Ev Idx": 5947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564855970.366, "dur": 4.055,
    "args": {
      "External id": 5949,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3054, "finished": true, "Ev Idx": 5948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564855971.720, "dur": 1.741,
    "args": {
      "External id": 5950,"Record function id": 0, "finished": true, "Ev Idx": 5949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564855972.475, "dur": 0.841,
    "args": {
      "External id": 5951,"Record function id": 0, "finished": true, "Ev Idx": 5950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564855974.825, "dur": 42.577,
    "args": {
      "External id": 5952,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3055, "finished": true, "Ev Idx": 5951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564855975.427, "dur": 2.136,
    "args": {
      "External id": 5953,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3055, "finished": true, "Ev Idx": 5952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564855975.825, "dur": 1.601,
    "args": {
      "External id": 5954,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3055, "finished": true, "Ev Idx": 5953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564855977.992, "dur": 37.184,
    "args": {
      "External id": 5955,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3056, "finished": true, "Ev Idx": 5954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564855988.680, "dur": 1.215,
    "args": {
      "External id": 5956,"Record function id": 0, "finished": true, "Ev Idx": 5955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564856015.810, "dur": 1.219,
    "args": {
      "External id": 5957,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3057, "finished": true, "Ev Idx": 5956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564856019.480, "dur": 42.164,
    "args": {
      "External id": 5958,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3058, "finished": true, "Ev Idx": 5957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564856069.612, "dur": 47.488,
    "args": {
      "External id": 5959,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3059, "finished": true, "Ev Idx": 5958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564856070.121, "dur": 4.458,
    "args": {
      "External id": 5960,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3059, "finished": true, "Ev Idx": 5959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564856072.036, "dur": 1.733,
    "args": {
      "External id": 5961,"Record function id": 0, "finished": true, "Ev Idx": 5960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856072.949, "dur": 0.681,
    "args": {
      "External id": 5962,"Record function id": 0, "finished": true, "Ev Idx": 5961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564856074.952, "dur": 42.048,
    "args": {
      "External id": 5963,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3060, "finished": true, "Ev Idx": 5962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856075.430, "dur": 1.772,
    "args": {
      "External id": 5964,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3060, "finished": true, "Ev Idx": 5963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856075.741, "dur": 1.334,
    "args": {
      "External id": 5965,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3060, "finished": true, "Ev Idx": 5964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564856077.531, "dur": 37.507,
    "args": {
      "External id": 5966,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3061, "finished": true, "Ev Idx": 5965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564856089.583, "dur": 1.078,
    "args": {
      "External id": 5967,"Record function id": 0, "finished": true, "Ev Idx": 5966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564856115.651, "dur": 1.032,
    "args": {
      "External id": 5968,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3062, "finished": true, "Ev Idx": 5967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564856120.270, "dur": 8.135,
    "args": {
      "External id": 5969,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3063, "finished": true, "Ev Idx": 5968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564856150.061, "dur": 0.289,
    "args": {
      "External id": 5970,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3064, "finished": true, "Ev Idx": 5969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564856152.552, "dur": 12.077,
    "args": {
      "External id": 5971,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3064, "finished": true, "Ev Idx": 5970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564856154.026, "dur": 0.630,
    "args": {
      "External id": 5972,"Record function id": 0, "finished": true, "Ev Idx": 5971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564856155.063, "dur": 0.078,
    "args": {
      "External id": 5973,"Record function id": 0, "finished": true, "Ev Idx": 5972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564856167.163, "dur": 12.344,
    "args": {
      "External id": 5974,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3065, "finished": true, "Ev Idx": 5973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564856183.460, "dur": 7.819,
    "args": {
      "External id": 5975,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3066, "finished": true, "Ev Idx": 5974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564856193.027, "dur": 9.057,
    "args": {
      "External id": 5976,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3067, "finished": true, "Ev Idx": 5975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564856203.753, "dur": 7.953,
    "args": {
      "External id": 5977,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3068, "finished": true, "Ev Idx": 5976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564856214.179, "dur": 0.128,
    "args": {
      "External id": 5978,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3069, "finished": true, "Ev Idx": 5977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564856215.184, "dur": 10.017,
    "args": {
      "External id": 5979,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3069, "finished": true, "Ev Idx": 5978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564856243.655, "dur": 38.667,
    "args": {
      "External id": 5980,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3070, "finished": true, "Ev Idx": 5979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564856244.046, "dur": 3.495,
    "args": {
      "External id": 5981,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3070, "finished": true, "Ev Idx": 5980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564856245.301, "dur": 1.366,
    "args": {
      "External id": 5982,"Record function id": 0, "finished": true, "Ev Idx": 5981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856246.002, "dur": 0.534,
    "args": {
      "External id": 5983,"Record function id": 0, "finished": true, "Ev Idx": 5982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564856248.085, "dur": 34.139,
    "args": {
      "External id": 5984,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3071, "finished": true, "Ev Idx": 5983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856248.590, "dur": 2.866,
    "args": {
      "External id": 5985,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3071, "finished": true, "Ev Idx": 5984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856248.934, "dur": 2.393,
    "args": {
      "External id": 5986,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3071, "finished": true, "Ev Idx": 5985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564856251.786, "dur": 28.421,
    "args": {
      "External id": 5987,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3072, "finished": true, "Ev Idx": 5986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564856262.466, "dur": 1.176,
    "args": {
      "External id": 5988,"Record function id": 0, "finished": true, "Ev Idx": 5987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564856280.774, "dur": 1.141,
    "args": {
      "External id": 5989,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3073, "finished": true, "Ev Idx": 5988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856285.608, "dur": 1.389,
    "args": {
      "External id": 5990,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3074, "finished": true, "Ev Idx": 5989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564856288.973, "dur": 2.598,
    "args": {
      "External id": 5991,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3075, "finished": true, "Ev Idx": 5990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856290.048, "dur": 0.697,
    "args": {
      "External id": 5992,"Record function id": 0, "finished": true, "Ev Idx": 5991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564856298.365, "dur": 40.923,
    "args": {
      "External id": 5993,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3076, "finished": true, "Ev Idx": 5992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564856298.665, "dur": 2.482,
    "args": {
      "External id": 5994,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3076, "finished": true, "Ev Idx": 5993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564856299.747, "dur": 0.957,
    "args": {
      "External id": 5995,"Record function id": 0, "finished": true, "Ev Idx": 5994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856300.271, "dur": 0.327,
    "args": {
      "External id": 5996,"Record function id": 0, "finished": true, "Ev Idx": 5995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564856301.481, "dur": 37.711,
    "args": {
      "External id": 5997,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3077, "finished": true, "Ev Idx": 5996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856301.826, "dur": 1.248,
    "args": {
      "External id": 5998,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3077, "finished": true, "Ev Idx": 5997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856302.075, "dur": 0.900,
    "args": {
      "External id": 5999,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3077, "finished": true, "Ev Idx": 5998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564856303.383, "dur": 33.850,
    "args": {
      "External id": 6000,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3078, "finished": true, "Ev Idx": 5999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564856314.076, "dur": 1.126,
    "args": {
      "External id": 6001,"Record function id": 0, "finished": true, "Ev Idx": 6000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564856337.906, "dur": 0.983,
    "args": {
      "External id": 6002,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3079, "finished": true, "Ev Idx": 6001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856341.412, "dur": 1.165,
    "args": {
      "External id": 6003,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3080, "finished": true, "Ev Idx": 6002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564856344.042, "dur": 1.894,
    "args": {
      "External id": 6004,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3081, "finished": true, "Ev Idx": 6003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856344.824, "dur": 0.456,
    "args": {
      "External id": 6005,"Record function id": 0, "finished": true, "Ev Idx": 6004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564856352.925, "dur": 36.055,
    "args": {
      "External id": 6006,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3082, "finished": true, "Ev Idx": 6005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564856353.187, "dur": 2.192,
    "args": {
      "External id": 6007,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3082, "finished": true, "Ev Idx": 6006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564856354.030, "dur": 0.935,
    "args": {
      "External id": 6008,"Record function id": 0, "finished": true, "Ev Idx": 6007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856354.598, "dur": 0.267,
    "args": {
      "External id": 6009,"Record function id": 0, "finished": true, "Ev Idx": 6008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564856355.666, "dur": 33.213,
    "args": {
      "External id": 6010,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3083, "finished": true, "Ev Idx": 6009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856356.081, "dur": 2.002,
    "args": {
      "External id": 6011,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3083, "finished": true, "Ev Idx": 6010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856356.386, "dur": 1.602,
    "args": {
      "External id": 6012,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3083, "finished": true, "Ev Idx": 6011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564856358.376, "dur": 28.795,
    "args": {
      "External id": 6013,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3084, "finished": true, "Ev Idx": 6012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564856366.940, "dur": 1.075,
    "args": {
      "External id": 6014,"Record function id": 0, "finished": true, "Ev Idx": 6013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564856387.693, "dur": 0.897,
    "args": {
      "External id": 6015,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3085, "finished": true, "Ev Idx": 6014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856391.316, "dur": 1.103,
    "args": {
      "External id": 6016,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3086, "finished": true, "Ev Idx": 6015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564856393.778, "dur": 1.600,
    "args": {
      "External id": 6017,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3087, "finished": true, "Ev Idx": 6016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856394.507, "dur": 0.373,
    "args": {
      "External id": 6018,"Record function id": 0, "finished": true, "Ev Idx": 6017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564856398.323, "dur": 2.141,
    "args": {
      "External id": 6019,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3088, "finished": true, "Ev Idx": 6018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856399.558, "dur": 0.346,
    "args": {
      "External id": 6020,"Record function id": 0, "finished": true, "Ev Idx": 6019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564856401.606, "dur": 1.139,
    "args": {
      "External id": 6021,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3088, "finished": true, "Ev Idx": 6020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856402.099, "dur": 0.334,
    "args": {
      "External id": 6022,"Record function id": 0, "finished": true, "Ev Idx": 6021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564856403.582, "dur": 10.427,
    "args": {
      "External id": 6023,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3088, "finished": true, "Ev Idx": 6022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856419.657, "dur": 4.163,
    "args": {
      "External id": 6024,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3089, "finished": true, "Ev Idx": 6023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856421.869, "dur": 0.586,
    "args": {
      "External id": 6025,"Record function id": 0, "finished": true, "Ev Idx": 6024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856426.010, "dur": 3.316,
    "args": {
      "External id": 6026,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3090, "finished": true, "Ev Idx": 6025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856427.617, "dur": 0.798,
    "args": {
      "External id": 6027,"Record function id": 0, "finished": true, "Ev Idx": 6026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564856431.184, "dur": 12.294,
    "args": {
      "External id": 6028,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3091, "finished": true, "Ev Idx": 6027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564856447.843, "dur": 16.678,
    "args": {
      "External id": 6029,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3092, "finished": true, "Ev Idx": 6028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564856467.394, "dur": 9.183,
    "args": {
      "External id": 6030,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3093, "finished": true, "Ev Idx": 6029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564856478.251, "dur": 7.877,
    "args": {
      "External id": 6031,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3094, "finished": true, "Ev Idx": 6030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564856488.199, "dur": 8.787,
    "args": {
      "External id": 6032,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3095, "finished": true, "Ev Idx": 6031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856499.579, "dur": 2.748,
    "args": {
      "External id": 6033,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3096, "finished": true, "Ev Idx": 6032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856501.180, "dur": 0.496,
    "args": {
      "External id": 6034,"Record function id": 0, "finished": true, "Ev Idx": 6033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856503.935, "dur": 1.775,
    "args": {
      "External id": 6035,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3097, "finished": true, "Ev Idx": 6034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856504.838, "dur": 0.339,
    "args": {
      "External id": 6036,"Record function id": 0, "finished": true, "Ev Idx": 6035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564856506.809, "dur": 8.144,
    "args": {
      "External id": 6037,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3098, "finished": true, "Ev Idx": 6036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564856516.847, "dur": 9.981,
    "args": {
      "External id": 6038,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3099, "finished": true, "Ev Idx": 6037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564856529.112, "dur": 6.727,
    "args": {
      "External id": 6039,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3100, "finished": true, "Ev Idx": 6038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564856537.088, "dur": 7.683,
    "args": {
      "External id": 6040,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3101, "finished": true, "Ev Idx": 6039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856568.858, "dur": 2.474,
    "args": {
      "External id": 6041,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3102, "finished": true, "Ev Idx": 6040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856570.114, "dur": 0.491,
    "args": {
      "External id": 6042,"Record function id": 0, "finished": true, "Ev Idx": 6041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856571.983, "dur": 2.687,
    "args": {
      "External id": 6043,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3103, "finished": true, "Ev Idx": 6042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856573.736, "dur": 0.465,
    "args": {
      "External id": 6044,"Record function id": 0, "finished": true, "Ev Idx": 6043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564856575.400, "dur": 2.660,
    "args": {
      "External id": 6045,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3104, "finished": true, "Ev Idx": 6044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856576.549, "dur": 0.838,
    "args": {
      "External id": 6046,"Record function id": 0, "finished": true, "Ev Idx": 6045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856578.781, "dur": 2.481,
    "args": {
      "External id": 6047,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3105, "finished": true, "Ev Idx": 6046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856580.264, "dur": 0.484,
    "args": {
      "External id": 6048,"Record function id": 0, "finished": true, "Ev Idx": 6047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856581.884, "dur": 1.543,
    "args": {
      "External id": 6049,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3106, "finished": true, "Ev Idx": 6048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856582.576, "dur": 0.409,
    "args": {
      "External id": 6050,"Record function id": 0, "finished": true, "Ev Idx": 6049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564856586.343, "dur": 3.929,
    "args": {
      "External id": 6051,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3107, "finished": true, "Ev Idx": 6050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856589.175, "dur": 0.387,
    "args": {
      "External id": 6052,"Record function id": 0, "finished": true, "Ev Idx": 6051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856592.782, "dur": 20.524,
    "args": {
      "External id": 6053,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3108, "finished": true, "Ev Idx": 6052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564856595.105, "dur": 15.151,
    "args": {
      "External id": 6054,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3108, "finished": true, "Ev Idx": 6053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564856597.273, "dur": 2.804,
    "args": {
      "External id": 6055,"Record function id": 0, "finished": true, "Ev Idx": 6054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564856598.083, "dur": 1.741,
    "args": {
      "External id": 6056,"Record function id": 0, "finished": true, "Ev Idx": 6055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564856600.819, "dur": 8.823,
    "args": {
      "External id": 6057,"Record function id": 0, "finished": true, "Ev Idx": 6056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564856610.817, "dur": 2.249,
    "args": {
      "External id": 6058,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3109, "finished": true, "Ev Idx": 6057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856616.522, "dur": 2.884,
    "args": {
      "External id": 6059,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3110, "finished": true, "Ev Idx": 6058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856618.375, "dur": 0.414,
    "args": {
      "External id": 6060,"Record function id": 0, "finished": true, "Ev Idx": 6059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856620.041, "dur": 2.761,
    "args": {
      "External id": 6061,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3111, "finished": true, "Ev Idx": 6060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856622.059, "dur": 0.289,
    "args": {
      "External id": 6062,"Record function id": 0, "finished": true, "Ev Idx": 6061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564856623.328, "dur": 2.170,
    "args": {
      "External id": 6063,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3112, "finished": true, "Ev Idx": 6062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856624.682, "dur": 0.357,
    "args": {
      "External id": 6064,"Record function id": 0, "finished": true, "Ev Idx": 6063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856626.191, "dur": 2.443,
    "args": {
      "External id": 6065,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3113, "finished": true, "Ev Idx": 6064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856627.817, "dur": 0.341,
    "args": {
      "External id": 6066,"Record function id": 0, "finished": true, "Ev Idx": 6065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856629.286, "dur": 2.079,
    "args": {
      "External id": 6067,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3114, "finished": true, "Ev Idx": 6066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856630.552, "dur": 0.353,
    "args": {
      "External id": 6068,"Record function id": 0, "finished": true, "Ev Idx": 6067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564856632.902, "dur": 1.537,
    "args": {
      "External id": 6069,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3115, "finished": true, "Ev Idx": 6068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856633.700, "dur": 0.333,
    "args": {
      "External id": 6070,"Record function id": 0, "finished": true, "Ev Idx": 6069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856635.900, "dur": 10.936,
    "args": {
      "External id": 6071,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3116, "finished": true, "Ev Idx": 6070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564856636.785, "dur": 8.600,
    "args": {
      "External id": 6072,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3116, "finished": true, "Ev Idx": 6071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564856637.496, "dur": 1.619,
    "args": {
      "External id": 6073,"Record function id": 0, "finished": true, "Ev Idx": 6072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564856637.806, "dur": 1.177,
    "args": {
      "External id": 6074,"Record function id": 0, "finished": true, "Ev Idx": 6073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564856639.402, "dur": 5.659,
    "args": {
      "External id": 6075,"Record function id": 0, "finished": true, "Ev Idx": 6074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564856645.758, "dur": 0.907,
    "args": {
      "External id": 6076,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3117, "finished": true, "Ev Idx": 6075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564856648.924, "dur": 2.861,
    "args": {
      "External id": 6077,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3118, "finished": true, "Ev Idx": 6076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856650.812, "dur": 0.425,
    "args": {
      "External id": 6078,"Record function id": 0, "finished": true, "Ev Idx": 6077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564856653.369, "dur": 45.001,
    "args": {
      "External id": 6079,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3119, "finished": true, "Ev Idx": 6078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564856654.020, "dur": 2.263,
    "args": {
      "External id": 6080,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3119, "finished": true, "Ev Idx": 6079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856655.522, "dur": 0.354,
    "args": {
      "External id": 6081,"Record function id": 0, "finished": true, "Ev Idx": 6080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856656.649, "dur": 3.519,
    "args": {
      "External id": 6082,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3120, "finished": true, "Ev Idx": 6081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564856658.370, "dur": 1.685,
    "args": {
      "External id": 6083,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3120, "finished": true, "Ev Idx": 6082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564856660.648, "dur": 1.333,
    "args": {
      "External id": 6084,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3121, "finished": true, "Ev Idx": 6083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856661.314, "dur": 0.277,
    "args": {
      "External id": 6085,"Record function id": 0, "finished": true, "Ev Idx": 6084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856662.266, "dur": 1.775,
    "args": {
      "External id": 6086,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3122, "finished": true, "Ev Idx": 6085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564856663.072, "dur": 0.863,
    "args": {
      "External id": 6087,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3122, "finished": true, "Ev Idx": 6086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564856664.695, "dur": 31.337,
    "args": {
      "External id": 6088,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3123, "finished": true, "Ev Idx": 6087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564856678.682, "dur": 0.999,
    "args": {
      "External id": 6089,"Record function id": 0, "finished": true, "Ev Idx": 6088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564856696.787, "dur": 1.155,
    "args": {
      "External id": 6090,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3124, "finished": true, "Ev Idx": 6089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564856701.634, "dur": 10.983,
    "args": {
      "External id": 6091,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3125, "finished": true, "Ev Idx": 6090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856715.665, "dur": 1.942,
    "args": {
      "External id": 6092,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3126, "finished": true, "Ev Idx": 6091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856716.574, "dur": 0.440,
    "args": {
      "External id": 6093,"Record function id": 0, "finished": true, "Ev Idx": 6092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856718.315, "dur": 1.258,
    "args": {
      "External id": 6094,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3126, "finished": true, "Ev Idx": 6093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856718.859, "dur": 0.307,
    "args": {
      "External id": 6095,"Record function id": 0, "finished": true, "Ev Idx": 6094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564856720.240, "dur": 0.994,
    "args": {
      "External id": 6096,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3126, "finished": true, "Ev Idx": 6095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856720.661, "dur": 0.261,
    "args": {
      "External id": 6097,"Record function id": 0, "finished": true, "Ev Idx": 6096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564856722.879, "dur": 7.401,
    "args": {
      "External id": 6098,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3126, "finished": true, "Ev Idx": 6097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564856735.647, "dur": 12.731,
    "args": {
      "External id": 6099,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3127, "finished": true, "Ev Idx": 6098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564856736.628, "dur": 0.380,
    "args": {
      "External id": 6100,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3127, "finished": true, "Ev Idx": 6099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564856737.683, "dur": 10.412,
    "args": {
      "External id": 6101,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3127, "finished": true, "Ev Idx": 6100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564856751.392, "dur": 0.212,
    "args": {
      "External id": 6102,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3128, "finished": true, "Ev Idx": 6101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564856763.806, "dur": 0.405,
    "args": {
      "External id": 6103,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3128, "finished": true, "Ev Idx": 6102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564856765.920, "dur": 39.836,
    "args": {
      "External id": 6104,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3128, "finished": true, "Ev Idx": 6103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564856766.362, "dur": 2.179,
    "args": {
      "External id": 6105,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3128, "finished": true, "Ev Idx": 6104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856767.558, "dur": 0.487,
    "args": {
      "External id": 6106,"Record function id": 0, "finished": true, "Ev Idx": 6105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856768.936, "dur": 2.656,
    "args": {
      "External id": 6107,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3129, "finished": true, "Ev Idx": 6106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856769.277, "dur": 2.190,
    "args": {
      "External id": 6108,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3129, "finished": true, "Ev Idx": 6107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564856772.053, "dur": 2.327,
    "args": {
      "External id": 6109,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3130, "finished": true, "Ev Idx": 6108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856773.682, "dur": 0.327,
    "args": {
      "External id": 6110,"Record function id": 0, "finished": true, "Ev Idx": 6109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856774.746, "dur": 1.978,
    "args": {
      "External id": 6111,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3131, "finished": true, "Ev Idx": 6110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856775.103, "dur": 1.544,
    "args": {
      "External id": 6112,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3131, "finished": true, "Ev Idx": 6111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564856777.146, "dur": 26.537,
    "args": {
      "External id": 6113,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3132, "finished": true, "Ev Idx": 6112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564856786.854, "dur": 1.177,
    "args": {
      "External id": 6114,"Record function id": 0, "finished": true, "Ev Idx": 6113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564856804.242, "dur": 1.190,
    "args": {
      "External id": 6115,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3133, "finished": true, "Ev Idx": 6114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564856807.740, "dur": 1.946,
    "args": {
      "External id": 6116,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3134, "finished": true, "Ev Idx": 6115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856808.785, "dur": 0.386,
    "args": {
      "External id": 6117,"Record function id": 0, "finished": true, "Ev Idx": 6116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564856812.245, "dur": 11.392,
    "args": {
      "External id": 6118,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3135, "finished": true, "Ev Idx": 6117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564856812.867, "dur": 10.621,
    "args": {
      "External id": 6119,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3135, "finished": true, "Ev Idx": 6118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564856814.078, "dur": 1.964,
    "args": {
      "External id": 6120,"Record function id": 0, "finished": true, "Ev Idx": 6119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564856814.522, "dur": 1.357,
    "args": {
      "External id": 6121,"Record function id": 0, "finished": true, "Ev Idx": 6120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564856816.342, "dur": 6.741,
    "args": {
      "External id": 6122,"Record function id": 0, "finished": true, "Ev Idx": 6121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856829.168, "dur": 1.703,
    "args": {
      "External id": 6123,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3136, "finished": true, "Ev Idx": 6122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856829.502, "dur": 1.280,
    "args": {
      "External id": 6124,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3136, "finished": true, "Ev Idx": 6123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564856843.167, "dur": 34.075,
    "args": {
      "External id": 6125,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3137, "finished": true, "Ev Idx": 6124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564856843.681, "dur": 2.872,
    "args": {
      "External id": 6126,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3137, "finished": true, "Ev Idx": 6125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564856845.074, "dur": 0.942,
    "args": {
      "External id": 6127,"Record function id": 0, "finished": true, "Ev Idx": 6126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564856845.526, "dur": 0.380,
    "args": {
      "External id": 6128,"Record function id": 0, "finished": true, "Ev Idx": 6127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564856846.921, "dur": 30.179,
    "args": {
      "External id": 6129,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3138, "finished": true, "Ev Idx": 6128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564856847.327, "dur": 1.678,
    "args": {
      "External id": 6130,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3138, "finished": true, "Ev Idx": 6129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564856847.599, "dur": 1.314,
    "args": {
      "External id": 6131,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3138, "finished": true, "Ev Idx": 6130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564856849.425, "dur": 25.639,
    "args": {
      "External id": 6132,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3139, "finished": true, "Ev Idx": 6131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564856859.809, "dur": 1.225,
    "args": {
      "External id": 6133,"Record function id": 0, "finished": true, "Ev Idx": 6132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564856875.778, "dur": 1.002,
    "args": {
      "External id": 6134,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3140, "finished": true, "Ev Idx": 6133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564856881.787, "dur": 9.546,
    "args": {
      "External id": 6135,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3141, "finished": true, "Ev Idx": 6134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564856900.415, "dur": 0.169,
    "args": {
      "External id": 6136,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3142, "finished": true, "Ev Idx": 6135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564856902.603, "dur": 10.624,
    "args": {
      "External id": 6137,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3142, "finished": true, "Ev Idx": 6136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564856904.236, "dur": 0.393,
    "args": {
      "External id": 6138,"Record function id": 0, "finished": true, "Ev Idx": 6137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564856905.157, "dur": 0.087,
    "args": {
      "External id": 6139,"Record function id": 0, "finished": true, "Ev Idx": 6138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564856915.971, "dur": 11.391,
    "args": {
      "External id": 6140,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3143, "finished": true, "Ev Idx": 6139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564856930.731, "dur": 8.329,
    "args": {
      "External id": 6141,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3144, "finished": true, "Ev Idx": 6140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564856941.073, "dur": 9.519,
    "args": {
      "External id": 6142,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3145, "finished": true, "Ev Idx": 6141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564856952.434, "dur": 6.988,
    "args": {
      "External id": 6143,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3146, "finished": true, "Ev Idx": 6142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564856961.931, "dur": 0.075,
    "args": {
      "External id": 6144,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3147, "finished": true, "Ev Idx": 6143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564856995.112, "dur": 64.821,
    "args": {
      "External id": 6145,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3147, "finished": true, "Ev Idx": 6144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564857077.567, "dur": 52.475,
    "args": {
      "External id": 6146,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3148, "finished": true, "Ev Idx": 6145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564857078.123, "dur": 4.779,
    "args": {
      "External id": 6147,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3148, "finished": true, "Ev Idx": 6146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564857080.398, "dur": 1.380,
    "args": {
      "External id": 6148,"Record function id": 0, "finished": true, "Ev Idx": 6147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857080.985, "dur": 0.579,
    "args": {
      "External id": 6149,"Record function id": 0, "finished": true, "Ev Idx": 6148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564857083.349, "dur": 46.558,
    "args": {
      "External id": 6150,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3149, "finished": true, "Ev Idx": 6149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857083.988, "dur": 1.895,
    "args": {
      "External id": 6151,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3149, "finished": true, "Ev Idx": 6150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857084.330, "dur": 1.413,
    "args": {
      "External id": 6152,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3149, "finished": true, "Ev Idx": 6151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564857086.209, "dur": 41.548,
    "args": {
      "External id": 6153,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3150, "finished": true, "Ev Idx": 6152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564857098.899, "dur": 1.254,
    "args": {
      "External id": 6154,"Record function id": 0, "finished": true, "Ev Idx": 6153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564857128.465, "dur": 1.088,
    "args": {
      "External id": 6155,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3151, "finished": true, "Ev Idx": 6154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564857138.394, "dur": 9.668,
    "args": {
      "External id": 6156,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3152, "finished": true, "Ev Idx": 6155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564857155.170, "dur": 81.001,
    "args": {
      "External id": 6157,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3153, "finished": true, "Ev Idx": 6156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564857155.471, "dur": 2.671,
    "args": {
      "External id": 6158,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3153, "finished": true, "Ev Idx": 6157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564857156.438, "dur": 1.078,
    "args": {
      "External id": 6159,"Record function id": 0, "finished": true, "Ev Idx": 6158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857156.907, "dur": 0.451,
    "args": {
      "External id": 6160,"Record function id": 0, "finished": true, "Ev Idx": 6159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564857158.473, "dur": 77.582,
    "args": {
      "External id": 6161,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3154, "finished": true, "Ev Idx": 6160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857158.882, "dur": 1.581,
    "args": {
      "External id": 6162,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3154, "finished": true, "Ev Idx": 6161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857159.195, "dur": 1.177,
    "args": {
      "External id": 6163,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3154, "finished": true, "Ev Idx": 6162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564857160.822, "dur": 73.077,
    "args": {
      "External id": 6164,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3155, "finished": true, "Ev Idx": 6163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564857208.902, "dur": 1.112,
    "args": {
      "External id": 6165,"Record function id": 0, "finished": true, "Ev Idx": 6164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564857234.510, "dur": 1.189,
    "args": {
      "External id": 6166,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3156, "finished": true, "Ev Idx": 6165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564857238.497, "dur": 10.540,
    "args": {
      "External id": 6167,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3157, "finished": true, "Ev Idx": 6166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564857256.357, "dur": 39.316,
    "args": {
      "External id": 6168,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3158, "finished": true, "Ev Idx": 6167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564857256.781, "dur": 3.184,
    "args": {
      "External id": 6169,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3158, "finished": true, "Ev Idx": 6168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564857258.355, "dur": 1.035,
    "args": {
      "External id": 6170,"Record function id": 0, "finished": true, "Ev Idx": 6169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857258.831, "dur": 0.415,
    "args": {
      "External id": 6171,"Record function id": 0, "finished": true, "Ev Idx": 6170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564857260.297, "dur": 35.273,
    "args": {
      "External id": 6172,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3159, "finished": true, "Ev Idx": 6171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857260.727, "dur": 1.494,
    "args": {
      "External id": 6173,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3159, "finished": true, "Ev Idx": 6172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857261.023, "dur": 1.106,
    "args": {
      "External id": 6174,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3159, "finished": true, "Ev Idx": 6173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564857262.550, "dur": 31.229,
    "args": {
      "External id": 6175,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3160, "finished": true, "Ev Idx": 6174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564857272.899, "dur": 1.158,
    "args": {
      "External id": 6176,"Record function id": 0, "finished": true, "Ev Idx": 6175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564857294.334, "dur": 0.956,
    "args": {
      "External id": 6177,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3161, "finished": true, "Ev Idx": 6176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564857298.465, "dur": 8.650,
    "args": {
      "External id": 6178,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3162, "finished": true, "Ev Idx": 6177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564857327.840, "dur": 0.199,
    "args": {
      "External id": 6179,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3163, "finished": true, "Ev Idx": 6178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564857330.034, "dur": 10.549,
    "args": {
      "External id": 6180,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3163, "finished": true, "Ev Idx": 6179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564857331.466, "dur": 0.493,
    "args": {
      "External id": 6181,"Record function id": 0, "finished": true, "Ev Idx": 6180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564857332.418, "dur": 0.078,
    "args": {
      "External id": 6182,"Record function id": 0, "finished": true, "Ev Idx": 6181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564857343.326, "dur": 12.637,
    "args": {
      "External id": 6183,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3164, "finished": true, "Ev Idx": 6182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564857359.626, "dur": 9.589,
    "args": {
      "External id": 6184,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3165, "finished": true, "Ev Idx": 6183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564857370.892, "dur": 9.453,
    "args": {
      "External id": 6185,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3166, "finished": true, "Ev Idx": 6184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564857381.915, "dur": 7.127,
    "args": {
      "External id": 6186,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3167, "finished": true, "Ev Idx": 6185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564857391.570, "dur": 0.080,
    "args": {
      "External id": 6187,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3168, "finished": true, "Ev Idx": 6186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564857392.526, "dur": 8.159,
    "args": {
      "External id": 6188,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3168, "finished": true, "Ev Idx": 6187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564857418.051, "dur": 73.145,
    "args": {
      "External id": 6189,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3169, "finished": true, "Ev Idx": 6188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564857418.423, "dur": 3.896,
    "args": {
      "External id": 6190,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3169, "finished": true, "Ev Idx": 6189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564857419.751, "dur": 1.352,
    "args": {
      "External id": 6191,"Record function id": 0, "finished": true, "Ev Idx": 6190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857420.502, "dur": 0.442,
    "args": {
      "External id": 6192,"Record function id": 0, "finished": true, "Ev Idx": 6191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564857422.638, "dur": 68.459,
    "args": {
      "External id": 6193,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3170, "finished": true, "Ev Idx": 6192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857423.218, "dur": 2.278,
    "args": {
      "External id": 6194,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3170, "finished": true, "Ev Idx": 6193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857423.542, "dur": 1.847,
    "args": {
      "External id": 6195,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3170, "finished": true, "Ev Idx": 6194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564857425.784, "dur": 62.978,
    "args": {
      "External id": 6196,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3171, "finished": true, "Ev Idx": 6195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564857471.104, "dur": 1.274,
    "args": {
      "External id": 6197,"Record function id": 0, "finished": true, "Ev Idx": 6196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564857489.412, "dur": 1.340,
    "args": {
      "External id": 6198,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3172, "finished": true, "Ev Idx": 6197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857494.837, "dur": 1.708,
    "args": {
      "External id": 6199,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3173, "finished": true, "Ev Idx": 6198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564857498.325, "dur": 3.060,
    "args": {
      "External id": 6200,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3174, "finished": true, "Ev Idx": 6199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857499.661, "dur": 0.710,
    "args": {
      "External id": 6201,"Record function id": 0, "finished": true, "Ev Idx": 6200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564857508.950, "dur": 43.674,
    "args": {
      "External id": 6202,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3175, "finished": true, "Ev Idx": 6201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564857509.371, "dur": 2.506,
    "args": {
      "External id": 6203,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3175, "finished": true, "Ev Idx": 6202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564857510.455, "dur": 0.907,
    "args": {
      "External id": 6204,"Record function id": 0, "finished": true, "Ev Idx": 6203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857510.941, "dur": 0.305,
    "args": {
      "External id": 6205,"Record function id": 0, "finished": true, "Ev Idx": 6204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564857512.262, "dur": 40.264,
    "args": {
      "External id": 6206,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3176, "finished": true, "Ev Idx": 6205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857512.826, "dur": 1.535,
    "args": {
      "External id": 6207,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3176, "finished": true, "Ev Idx": 6206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857513.233, "dur": 1.026,
    "args": {
      "External id": 6208,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3176, "finished": true, "Ev Idx": 6207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564857514.815, "dur": 35.788,
    "args": {
      "External id": 6209,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3177, "finished": true, "Ev Idx": 6208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564857526.912, "dur": 1.097,
    "args": {
      "External id": 6210,"Record function id": 0, "finished": true, "Ev Idx": 6209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564857551.160, "dur": 1.066,
    "args": {
      "External id": 6211,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3178, "finished": true, "Ev Idx": 6210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857554.840, "dur": 1.283,
    "args": {
      "External id": 6212,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3179, "finished": true, "Ev Idx": 6211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564857557.440, "dur": 1.677,
    "args": {
      "External id": 6213,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3180, "finished": true, "Ev Idx": 6212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857558.195, "dur": 0.409,
    "args": {
      "External id": 6214,"Record function id": 0, "finished": true, "Ev Idx": 6213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564857565.693, "dur": 38.727,
    "args": {
      "External id": 6215,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3181, "finished": true, "Ev Idx": 6214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564857566.020, "dur": 2.142,
    "args": {
      "External id": 6216,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3181, "finished": true, "Ev Idx": 6215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564857566.886, "dur": 0.811,
    "args": {
      "External id": 6217,"Record function id": 0, "finished": true, "Ev Idx": 6216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857567.287, "dur": 0.276,
    "args": {
      "External id": 6218,"Record function id": 0, "finished": true, "Ev Idx": 6217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564857568.463, "dur": 35.858,
    "args": {
      "External id": 6219,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3182, "finished": true, "Ev Idx": 6218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857568.856, "dur": 2.031,
    "args": {
      "External id": 6220,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3182, "finished": true, "Ev Idx": 6219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857569.227, "dur": 1.544,
    "args": {
      "External id": 6221,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3182, "finished": true, "Ev Idx": 6220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564857571.188, "dur": 31.191,
    "args": {
      "External id": 6222,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3183, "finished": true, "Ev Idx": 6221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564857582.307, "dur": 1.030,
    "args": {
      "External id": 6223,"Record function id": 0, "finished": true, "Ev Idx": 6222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564857602.927, "dur": 1.090,
    "args": {
      "External id": 6224,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3184, "finished": true, "Ev Idx": 6223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857606.769, "dur": 1.327,
    "args": {
      "External id": 6225,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3185, "finished": true, "Ev Idx": 6224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564857609.500, "dur": 1.681,
    "args": {
      "External id": 6226,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3186, "finished": true, "Ev Idx": 6225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857610.363, "dur": 0.351,
    "args": {
      "External id": 6227,"Record function id": 0, "finished": true, "Ev Idx": 6226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564857613.995, "dur": 2.491,
    "args": {
      "External id": 6228,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3187, "finished": true, "Ev Idx": 6227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857615.434, "dur": 0.529,
    "args": {
      "External id": 6229,"Record function id": 0, "finished": true, "Ev Idx": 6228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564857617.624, "dur": 1.141,
    "args": {
      "External id": 6230,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3187, "finished": true, "Ev Idx": 6229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857618.125, "dur": 0.289,
    "args": {
      "External id": 6231,"Record function id": 0, "finished": true, "Ev Idx": 6230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564857619.664, "dur": 10.819,
    "args": {
      "External id": 6232,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3187, "finished": true, "Ev Idx": 6231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857636.074, "dur": 5.074,
    "args": {
      "External id": 6233,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3188, "finished": true, "Ev Idx": 6232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857639.064, "dur": 0.690,
    "args": {
      "External id": 6234,"Record function id": 0, "finished": true, "Ev Idx": 6233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857643.567, "dur": 2.102,
    "args": {
      "External id": 6235,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3189, "finished": true, "Ev Idx": 6234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857644.636, "dur": 0.472,
    "args": {
      "External id": 6236,"Record function id": 0, "finished": true, "Ev Idx": 6235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564857647.709, "dur": 10.809,
    "args": {
      "External id": 6237,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3190, "finished": true, "Ev Idx": 6236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564857662.565, "dur": 18.454,
    "args": {
      "External id": 6238,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3191, "finished": true, "Ev Idx": 6237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564857684.184, "dur": 9.387,
    "args": {
      "External id": 6239,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3192, "finished": true, "Ev Idx": 6238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564857695.153, "dur": 8.411,
    "args": {
      "External id": 6240,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3193, "finished": true, "Ev Idx": 6239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564857705.584, "dur": 7.583,
    "args": {
      "External id": 6241,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3194, "finished": true, "Ev Idx": 6240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857715.503, "dur": 3.020,
    "args": {
      "External id": 6242,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3195, "finished": true, "Ev Idx": 6241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857717.373, "dur": 0.447,
    "args": {
      "External id": 6243,"Record function id": 0, "finished": true, "Ev Idx": 6242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857720.103, "dur": 1.979,
    "args": {
      "External id": 6244,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3196, "finished": true, "Ev Idx": 6243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857721.178, "dur": 0.362,
    "args": {
      "External id": 6245,"Record function id": 0, "finished": true, "Ev Idx": 6244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564857723.136, "dur": 6.911,
    "args": {
      "External id": 6246,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3197, "finished": true, "Ev Idx": 6245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564857731.768, "dur": 11.516,
    "args": {
      "External id": 6247,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3198, "finished": true, "Ev Idx": 6246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564857745.603, "dur": 6.847,
    "args": {
      "External id": 6248,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3199, "finished": true, "Ev Idx": 6247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564857753.908, "dur": 7.167,
    "args": {
      "External id": 6249,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3200, "finished": true, "Ev Idx": 6248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857785.037, "dur": 3.334,
    "args": {
      "External id": 6250,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3201, "finished": true, "Ev Idx": 6249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857787.162, "dur": 0.491,
    "args": {
      "External id": 6251,"Record function id": 0, "finished": true, "Ev Idx": 6250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857789.016, "dur": 2.445,
    "args": {
      "External id": 6252,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3202, "finished": true, "Ev Idx": 6251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857790.690, "dur": 0.311,
    "args": {
      "External id": 6253,"Record function id": 0, "finished": true, "Ev Idx": 6252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564857792.197, "dur": 2.147,
    "args": {
      "External id": 6254,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3203, "finished": true, "Ev Idx": 6253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857793.078, "dur": 0.627,
    "args": {
      "External id": 6255,"Record function id": 0, "finished": true, "Ev Idx": 6254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857795.093, "dur": 1.870,
    "args": {
      "External id": 6256,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3204, "finished": true, "Ev Idx": 6255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857796.056, "dur": 0.384,
    "args": {
      "External id": 6257,"Record function id": 0, "finished": true, "Ev Idx": 6256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857797.579, "dur": 1.470,
    "args": {
      "External id": 6258,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3205, "finished": true, "Ev Idx": 6257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857798.310, "dur": 0.290,
    "args": {
      "External id": 6259,"Record function id": 0, "finished": true, "Ev Idx": 6258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564857802.014, "dur": 2.910,
    "args": {
      "External id": 6260,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3206, "finished": true, "Ev Idx": 6259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857803.844, "dur": 0.363,
    "args": {
      "External id": 6261,"Record function id": 0, "finished": true, "Ev Idx": 6260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857807.289, "dur": 22.338,
    "args": {
      "External id": 6262,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3207, "finished": true, "Ev Idx": 6261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564857809.683, "dur": 17.792,
    "args": {
      "External id": 6263,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3207, "finished": true, "Ev Idx": 6262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564857812.224, "dur": 3.210,
    "args": {
      "External id": 6264,"Record function id": 0, "finished": true, "Ev Idx": 6263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564857813.344, "dur": 1.865,
    "args": {
      "External id": 6265,"Record function id": 0, "finished": true, "Ev Idx": 6264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564857816.096, "dur": 10.809,
    "args": {
      "External id": 6266,"Record function id": 0, "finished": true, "Ev Idx": 6265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564857827.995, "dur": 1.407,
    "args": {
      "External id": 6267,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3208, "finished": true, "Ev Idx": 6266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857832.919, "dur": 1.818,
    "args": {
      "External id": 6268,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3209, "finished": true, "Ev Idx": 6267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857833.810, "dur": 0.380,
    "args": {
      "External id": 6269,"Record function id": 0, "finished": true, "Ev Idx": 6268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857835.393, "dur": 1.656,
    "args": {
      "External id": 6270,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3210, "finished": true, "Ev Idx": 6269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857836.299, "dur": 0.299,
    "args": {
      "External id": 6271,"Record function id": 0, "finished": true, "Ev Idx": 6270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564857837.570, "dur": 3.095,
    "args": {
      "External id": 6272,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3211, "finished": true, "Ev Idx": 6271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857839.631, "dur": 0.626,
    "args": {
      "External id": 6273,"Record function id": 0, "finished": true, "Ev Idx": 6272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857841.267, "dur": 3.727,
    "args": {
      "External id": 6274,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3212, "finished": true, "Ev Idx": 6273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857844.072, "dur": 0.449,
    "args": {
      "External id": 6275,"Record function id": 0, "finished": true, "Ev Idx": 6274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857845.612, "dur": 2.026,
    "args": {
      "External id": 6276,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3213, "finished": true, "Ev Idx": 6275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857846.928, "dur": 0.272,
    "args": {
      "External id": 6277,"Record function id": 0, "finished": true, "Ev Idx": 6276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564857849.310, "dur": 1.721,
    "args": {
      "External id": 6278,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3214, "finished": true, "Ev Idx": 6277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857850.272, "dur": 0.317,
    "args": {
      "External id": 6279,"Record function id": 0, "finished": true, "Ev Idx": 6278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857852.576, "dur": 10.638,
    "args": {
      "External id": 6280,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3215, "finished": true, "Ev Idx": 6279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564857853.472, "dur": 8.258,
    "args": {
      "External id": 6281,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3215, "finished": true, "Ev Idx": 6280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564857854.114, "dur": 1.584,
    "args": {
      "External id": 6282,"Record function id": 0, "finished": true, "Ev Idx": 6281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564857854.441, "dur": 1.135,
    "args": {
      "External id": 6283,"Record function id": 0, "finished": true, "Ev Idx": 6282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564857855.992, "dur": 5.429,
    "args": {
      "External id": 6284,"Record function id": 0, "finished": true, "Ev Idx": 6283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564857862.147, "dur": 0.862,
    "args": {
      "External id": 6285,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3216, "finished": true, "Ev Idx": 6284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564857865.280, "dur": 2.234,
    "args": {
      "External id": 6286,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3217, "finished": true, "Ev Idx": 6285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857866.529, "dur": 0.500,
    "args": {
      "External id": 6287,"Record function id": 0, "finished": true, "Ev Idx": 6286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564857869.329, "dur": 44.919,
    "args": {
      "External id": 6288,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3218, "finished": true, "Ev Idx": 6287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564857870.160, "dur": 2.387,
    "args": {
      "External id": 6289,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3218, "finished": true, "Ev Idx": 6288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857871.811, "dur": 0.312,
    "args": {
      "External id": 6290,"Record function id": 0, "finished": true, "Ev Idx": 6289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857873.060, "dur": 4.094,
    "args": {
      "External id": 6291,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3219, "finished": true, "Ev Idx": 6290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564857874.800, "dur": 2.187,
    "args": {
      "External id": 6292,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3219, "finished": true, "Ev Idx": 6291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564857877.675, "dur": 1.330,
    "args": {
      "External id": 6293,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3220, "finished": true, "Ev Idx": 6292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857878.349, "dur": 0.262,
    "args": {
      "External id": 6294,"Record function id": 0, "finished": true, "Ev Idx": 6293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857879.304, "dur": 1.779,
    "args": {
      "External id": 6295,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3221, "finished": true, "Ev Idx": 6294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564857880.116, "dur": 0.872,
    "args": {
      "External id": 6296,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3221, "finished": true, "Ev Idx": 6295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564857881.772, "dur": 30.412,
    "args": {
      "External id": 6297,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3222, "finished": true, "Ev Idx": 6296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564857894.539, "dur": 0.998,
    "args": {
      "External id": 6298,"Record function id": 0, "finished": true, "Ev Idx": 6297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564857912.749, "dur": 1.143,
    "args": {
      "External id": 6299,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3223, "finished": true, "Ev Idx": 6298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564857917.336, "dur": 10.016,
    "args": {
      "External id": 6300,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3224, "finished": true, "Ev Idx": 6299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857930.455, "dur": 1.917,
    "args": {
      "External id": 6301,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3225, "finished": true, "Ev Idx": 6300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857931.240, "dur": 0.488,
    "args": {
      "External id": 6302,"Record function id": 0, "finished": true, "Ev Idx": 6301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857933.029, "dur": 1.155,
    "args": {
      "External id": 6303,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3225, "finished": true, "Ev Idx": 6302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857933.524, "dur": 0.285,
    "args": {
      "External id": 6304,"Record function id": 0, "finished": true, "Ev Idx": 6303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564857934.841, "dur": 1.200,
    "args": {
      "External id": 6305,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3225, "finished": true, "Ev Idx": 6304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857935.437, "dur": 0.292,
    "args": {
      "External id": 6306,"Record function id": 0, "finished": true, "Ev Idx": 6305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564857937.719, "dur": 7.484,
    "args": {
      "External id": 6307,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3225, "finished": true, "Ev Idx": 6306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564857950.632, "dur": 14.222,
    "args": {
      "External id": 6308,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3226, "finished": true, "Ev Idx": 6307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564857951.568, "dur": 0.428,
    "args": {
      "External id": 6309,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3226, "finished": true, "Ev Idx": 6308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564857952.646, "dur": 11.942,
    "args": {
      "External id": 6310,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3226, "finished": true, "Ev Idx": 6309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564857967.766, "dur": 0.209,
    "args": {
      "External id": 6311,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3227, "finished": true, "Ev Idx": 6310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564857978.457, "dur": 0.614,
    "args": {
      "External id": 6312,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3227, "finished": true, "Ev Idx": 6311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564857980.974, "dur": 42.873,
    "args": {
      "External id": 6313,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3227, "finished": true, "Ev Idx": 6312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564857981.664, "dur": 2.079,
    "args": {
      "External id": 6314,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3227, "finished": true, "Ev Idx": 6313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857982.666, "dur": 0.550,
    "args": {
      "External id": 6315,"Record function id": 0, "finished": true, "Ev Idx": 6314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857984.123, "dur": 2.578,
    "args": {
      "External id": 6316,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3228, "finished": true, "Ev Idx": 6315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857984.490, "dur": 2.081,
    "args": {
      "External id": 6317,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3228, "finished": true, "Ev Idx": 6316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564857987.156, "dur": 2.137,
    "args": {
      "External id": 6318,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3229, "finished": true, "Ev Idx": 6317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564857988.387, "dur": 0.555,
    "args": {
      "External id": 6319,"Record function id": 0, "finished": true, "Ev Idx": 6318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564857989.588, "dur": 2.962,
    "args": {
      "External id": 6320,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3230, "finished": true, "Ev Idx": 6319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564857989.994, "dur": 2.492,
    "args": {
      "External id": 6321,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3230, "finished": true, "Ev Idx": 6320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564857992.949, "dur": 28.960,
    "args": {
      "External id": 6322,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3231, "finished": true, "Ev Idx": 6321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564858003.721, "dur": 1.141,
    "args": {
      "External id": 6323,"Record function id": 0, "finished": true, "Ev Idx": 6322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564858022.466, "dur": 1.100,
    "args": {
      "External id": 6324,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3232, "finished": true, "Ev Idx": 6323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858025.888, "dur": 1.795,
    "args": {
      "External id": 6325,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3233, "finished": true, "Ev Idx": 6324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858026.807, "dur": 0.410,
    "args": {
      "External id": 6326,"Record function id": 0, "finished": true, "Ev Idx": 6325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564858030.289, "dur": 12.510,
    "args": {
      "External id": 6327,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3234, "finished": true, "Ev Idx": 6326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564858030.800, "dur": 11.866,
    "args": {
      "External id": 6328,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3234, "finished": true, "Ev Idx": 6327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564858032.130, "dur": 2.916,
    "args": {
      "External id": 6329,"Record function id": 0, "finished": true, "Ev Idx": 6328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564858032.533, "dur": 2.367,
    "args": {
      "External id": 6330,"Record function id": 0, "finished": true, "Ev Idx": 6329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564858035.376, "dur": 6.789,
    "args": {
      "External id": 6331,"Record function id": 0, "finished": true, "Ev Idx": 6330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564858048.708, "dur": 2.558,
    "args": {
      "External id": 6332,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3235, "finished": true, "Ev Idx": 6331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858049.046, "dur": 2.111,
    "args": {
      "External id": 6333,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3235, "finished": true, "Ev Idx": 6332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564858065.088, "dur": 66.769,
    "args": {
      "External id": 6334,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3236, "finished": true, "Ev Idx": 6333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564858065.604, "dur": 24.075,
    "args": {
      "External id": 6335,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3236, "finished": true, "Ev Idx": 6334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858068.348, "dur": 20.061,
    "args": {
      "External id": 6336,"Record function id": 0, "finished": true, "Ev Idx": 6335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858086.877, "dur": 1.051,
    "args": {
      "External id": 6337,"Record function id": 0, "finished": true, "Ev Idx": 6336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564858090.486, "dur": 41.220,
    "args": {
      "External id": 6338,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3237, "finished": true, "Ev Idx": 6337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564858091.138, "dur": 2.924,
    "args": {
      "External id": 6339,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3237, "finished": true, "Ev Idx": 6338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858091.734, "dur": 2.249,
    "args": {
      "External id": 6340,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3237, "finished": true, "Ev Idx": 6339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564858094.723, "dur": 34.415,
    "args": {
      "External id": 6341,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3238, "finished": true, "Ev Idx": 6340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564858110.627, "dur": 1.471,
    "args": {
      "External id": 6342,"Record function id": 0, "finished": true, "Ev Idx": 6341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564858130.040, "dur": 1.218,
    "args": {
      "External id": 6343,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3239, "finished": true, "Ev Idx": 6342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564858138.246, "dur": 10.197,
    "args": {
      "External id": 6344,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3240, "finished": true, "Ev Idx": 6343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564858159.733, "dur": 0.190,
    "args": {
      "External id": 6345,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3241, "finished": true, "Ev Idx": 6344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564858162.354, "dur": 13.197,
    "args": {
      "External id": 6346,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3241, "finished": true, "Ev Idx": 6345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564858164.394, "dur": 0.445,
    "args": {
      "External id": 6347,"Record function id": 0, "finished": true, "Ev Idx": 6346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564858165.216, "dur": 0.244,
    "args": {
      "External id": 6348,"Record function id": 0, "finished": true, "Ev Idx": 6347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564858178.611, "dur": 14.995,
    "args": {
      "External id": 6349,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3242, "finished": true, "Ev Idx": 6348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564858197.919, "dur": 9.900,
    "args": {
      "External id": 6350,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3243, "finished": true, "Ev Idx": 6349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564858210.037, "dur": 9.423,
    "args": {
      "External id": 6351,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3244, "finished": true, "Ev Idx": 6350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564858221.338, "dur": 9.844,
    "args": {
      "External id": 6352,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3245, "finished": true, "Ev Idx": 6351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564858234.217, "dur": 0.105,
    "args": {
      "External id": 6353,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3246, "finished": true, "Ev Idx": 6352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564858235.402, "dur": 9.226,
    "args": {
      "External id": 6354,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3246, "finished": true, "Ev Idx": 6353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564858259.019, "dur": 117.851,
    "args": {
      "External id": 6355,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3247, "finished": true, "Ev Idx": 6354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564858259.683, "dur": 4.422,
    "args": {
      "External id": 6356,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3247, "finished": true, "Ev Idx": 6355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858261.898, "dur": 1.415,
    "args": {
      "External id": 6357,"Record function id": 0, "finished": true, "Ev Idx": 6356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858262.717, "dur": 0.467,
    "args": {
      "External id": 6358,"Record function id": 0, "finished": true, "Ev Idx": 6357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564858264.490, "dur": 112.251,
    "args": {
      "External id": 6359,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3248, "finished": true, "Ev Idx": 6358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564858265.183, "dur": 1.622,
    "args": {
      "External id": 6360,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3248, "finished": true, "Ev Idx": 6359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858265.519, "dur": 1.192,
    "args": {
      "External id": 6361,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3248, "finished": true, "Ev Idx": 6360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564858267.147, "dur": 107.114,
    "args": {
      "External id": 6362,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3249, "finished": true, "Ev Idx": 6361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564858342.811, "dur": 1.261,
    "args": {
      "External id": 6363,"Record function id": 0, "finished": true, "Ev Idx": 6362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564858374.970, "dur": 1.337,
    "args": {
      "External id": 6364,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3250, "finished": true, "Ev Idx": 6363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564858386.853, "dur": 13.250,
    "args": {
      "External id": 6365,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3251, "finished": true, "Ev Idx": 6364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564858409.060, "dur": 75.915,
    "args": {
      "External id": 6366,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3252, "finished": true, "Ev Idx": 6365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564858409.423, "dur": 3.375,
    "args": {
      "External id": 6367,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3252, "finished": true, "Ev Idx": 6366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858410.789, "dur": 1.253,
    "args": {
      "External id": 6368,"Record function id": 0, "finished": true, "Ev Idx": 6367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858411.393, "dur": 0.523,
    "args": {
      "External id": 6369,"Record function id": 0, "finished": true, "Ev Idx": 6368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564858413.171, "dur": 71.710,
    "args": {
      "External id": 6370,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3253, "finished": true, "Ev Idx": 6369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564858413.621, "dur": 1.640,
    "args": {
      "External id": 6371,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3253, "finished": true, "Ev Idx": 6370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858413.918, "dur": 1.242,
    "args": {
      "External id": 6372,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3253, "finished": true, "Ev Idx": 6371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564858415.640, "dur": 66.992,
    "args": {
      "External id": 6373,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3254, "finished": true, "Ev Idx": 6372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564858458.632, "dur": 1.217,
    "args": {
      "External id": 6374,"Record function id": 0, "finished": true, "Ev Idx": 6373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564858483.269, "dur": 1.243,
    "args": {
      "External id": 6375,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3255, "finished": true, "Ev Idx": 6374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564858487.511, "dur": 9.456,
    "args": {
      "External id": 6376,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3256, "finished": true, "Ev Idx": 6375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564858503.989, "dur": 40.656,
    "args": {
      "External id": 6377,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3257, "finished": true, "Ev Idx": 6376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564858504.524, "dur": 3.086,
    "args": {
      "External id": 6378,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3257, "finished": true, "Ev Idx": 6377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858505.711, "dur": 1.320,
    "args": {
      "External id": 6379,"Record function id": 0, "finished": true, "Ev Idx": 6378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858506.343, "dur": 0.576,
    "args": {
      "External id": 6380,"Record function id": 0, "finished": true, "Ev Idx": 6379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564858507.949, "dur": 36.594,
    "args": {
      "External id": 6381,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3258, "finished": true, "Ev Idx": 6380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564858508.335, "dur": 1.867,
    "args": {
      "External id": 6382,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3258, "finished": true, "Ev Idx": 6381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858508.680, "dur": 1.429,
    "args": {
      "External id": 6383,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3258, "finished": true, "Ev Idx": 6382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564858510.545, "dur": 32.080,
    "args": {
      "External id": 6384,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3259, "finished": true, "Ev Idx": 6383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564858521.511, "dur": 1.137,
    "args": {
      "External id": 6385,"Record function id": 0, "finished": true, "Ev Idx": 6384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564858543.206, "dur": 1.064,
    "args": {
      "External id": 6386,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3260, "finished": true, "Ev Idx": 6385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564858547.346, "dur": 9.569,
    "args": {
      "External id": 6387,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3261, "finished": true, "Ev Idx": 6386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564858577.983, "dur": 0.373,
    "args": {
      "External id": 6388,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3262, "finished": true, "Ev Idx": 6387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564858580.663, "dur": 10.071,
    "args": {
      "External id": 6389,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3262, "finished": true, "Ev Idx": 6388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564858582.130, "dur": 0.682,
    "args": {
      "External id": 6390,"Record function id": 0, "finished": true, "Ev Idx": 6389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564858583.121, "dur": 0.283,
    "args": {
      "External id": 6391,"Record function id": 0, "finished": true, "Ev Idx": 6390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564858593.264, "dur": 12.675,
    "args": {
      "External id": 6392,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3263, "finished": true, "Ev Idx": 6391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564858609.546, "dur": 9.270,
    "args": {
      "External id": 6393,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3264, "finished": true, "Ev Idx": 6392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564858620.711, "dur": 8.898,
    "args": {
      "External id": 6394,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3265, "finished": true, "Ev Idx": 6393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564858631.357, "dur": 8.467,
    "args": {
      "External id": 6395,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3266, "finished": true, "Ev Idx": 6394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564858642.774, "dur": 0.129,
    "args": {
      "External id": 6396,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3267, "finished": true, "Ev Idx": 6395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564858643.774, "dur": 8.539,
    "args": {
      "External id": 6397,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3267, "finished": true, "Ev Idx": 6396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564858671.134, "dur": 35.535,
    "args": {
      "External id": 6398,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3268, "finished": true, "Ev Idx": 6397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564858671.979, "dur": 4.010,
    "args": {
      "External id": 6399,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3268, "finished": true, "Ev Idx": 6398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858673.661, "dur": 1.506,
    "args": {
      "External id": 6400,"Record function id": 0, "finished": true, "Ev Idx": 6399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858674.468, "dur": 0.587,
    "args": {
      "External id": 6401,"Record function id": 0, "finished": true, "Ev Idx": 6400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564858676.377, "dur": 30.173,
    "args": {
      "External id": 6402,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3269, "finished": true, "Ev Idx": 6401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564858676.947, "dur": 1.605,
    "args": {
      "External id": 6403,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3269, "finished": true, "Ev Idx": 6402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858677.329, "dur": 1.083,
    "args": {
      "External id": 6404,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3269, "finished": true, "Ev Idx": 6403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564858678.906, "dur": 25.711,
    "args": {
      "External id": 6405,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3270, "finished": true, "Ev Idx": 6404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564858688.833, "dur": 1.017,
    "args": {
      "External id": 6406,"Record function id": 0, "finished": true, "Ev Idx": 6405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564858705.170, "dur": 1.085,
    "args": {
      "External id": 6407,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3271, "finished": true, "Ev Idx": 6406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858710.026, "dur": 1.586,
    "args": {
      "External id": 6408,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3272, "finished": true, "Ev Idx": 6407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858713.454, "dur": 2.510,
    "args": {
      "External id": 6409,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3273, "finished": true, "Ev Idx": 6408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858714.486, "dur": 0.645,
    "args": {
      "External id": 6410,"Record function id": 0, "finished": true, "Ev Idx": 6409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564858722.447, "dur": 39.579,
    "args": {
      "External id": 6411,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3274, "finished": true, "Ev Idx": 6410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564858722.788, "dur": 2.250,
    "args": {
      "External id": 6412,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3274, "finished": true, "Ev Idx": 6411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858723.738, "dur": 0.859,
    "args": {
      "External id": 6413,"Record function id": 0, "finished": true, "Ev Idx": 6412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858724.237, "dur": 0.270,
    "args": {
      "External id": 6414,"Record function id": 0, "finished": true, "Ev Idx": 6413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564858725.342, "dur": 36.605,
    "args": {
      "External id": 6415,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3275, "finished": true, "Ev Idx": 6414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564858725.713, "dur": 1.515,
    "args": {
      "External id": 6416,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3275, "finished": true, "Ev Idx": 6415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858725.998, "dur": 1.129,
    "args": {
      "External id": 6417,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3275, "finished": true, "Ev Idx": 6416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564858727.519, "dur": 32.571,
    "args": {
      "External id": 6418,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3276, "finished": true, "Ev Idx": 6417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564858736.487, "dur": 1.111,
    "args": {
      "External id": 6419,"Record function id": 0, "finished": true, "Ev Idx": 6418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564858760.638, "dur": 1.006,
    "args": {
      "External id": 6420,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3277, "finished": true, "Ev Idx": 6419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858764.314, "dur": 1.254,
    "args": {
      "External id": 6421,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3278, "finished": true, "Ev Idx": 6420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858767.363, "dur": 2.231,
    "args": {
      "External id": 6422,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3279, "finished": true, "Ev Idx": 6421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858768.417, "dur": 0.532,
    "args": {
      "External id": 6423,"Record function id": 0, "finished": true, "Ev Idx": 6422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564858776.697, "dur": 36.221,
    "args": {
      "External id": 6424,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3280, "finished": true, "Ev Idx": 6423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564858776.966, "dur": 2.026,
    "args": {
      "External id": 6425,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3280, "finished": true, "Ev Idx": 6424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858777.731, "dur": 0.791,
    "args": {
      "External id": 6426,"Record function id": 0, "finished": true, "Ev Idx": 6425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858778.159, "dur": 0.260,
    "args": {
      "External id": 6427,"Record function id": 0, "finished": true, "Ev Idx": 6426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564858779.287, "dur": 33.534,
    "args": {
      "External id": 6428,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3281, "finished": true, "Ev Idx": 6427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564858779.651, "dur": 1.225,
    "args": {
      "External id": 6429,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3281, "finished": true, "Ev Idx": 6428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858779.898, "dur": 0.882,
    "args": {
      "External id": 6430,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3281, "finished": true, "Ev Idx": 6429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564858781.171, "dur": 29.891,
    "args": {
      "External id": 6431,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3282, "finished": true, "Ev Idx": 6430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564858791.242, "dur": 1.093,
    "args": {
      "External id": 6432,"Record function id": 0, "finished": true, "Ev Idx": 6431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564858811.589, "dur": 0.947,
    "args": {
      "External id": 6433,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3283, "finished": true, "Ev Idx": 6432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564858815.366, "dur": 1.120,
    "args": {
      "External id": 6434,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3284, "finished": true, "Ev Idx": 6433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564858817.876, "dur": 1.563,
    "args": {
      "External id": 6435,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3285, "finished": true, "Ev Idx": 6434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858818.548, "dur": 0.376,
    "args": {
      "External id": 6436,"Record function id": 0, "finished": true, "Ev Idx": 6435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564858822.311, "dur": 2.262,
    "args": {
      "External id": 6437,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3286, "finished": true, "Ev Idx": 6436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858823.738, "dur": 0.318,
    "args": {
      "External id": 6438,"Record function id": 0, "finished": true, "Ev Idx": 6437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564858825.787, "dur": 1.079,
    "args": {
      "External id": 6439,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3286, "finished": true, "Ev Idx": 6438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858826.234, "dur": 0.288,
    "args": {
      "External id": 6440,"Record function id": 0, "finished": true, "Ev Idx": 6439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564858827.661, "dur": 10.035,
    "args": {
      "External id": 6441,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3286, "finished": true, "Ev Idx": 6440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564858843.845, "dur": 4.682,
    "args": {
      "External id": 6442,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3287, "finished": true, "Ev Idx": 6441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858846.414, "dur": 0.640,
    "args": {
      "External id": 6443,"Record function id": 0, "finished": true, "Ev Idx": 6442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564858850.810, "dur": 2.924,
    "args": {
      "External id": 6444,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3288, "finished": true, "Ev Idx": 6443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858852.543, "dur": 0.555,
    "args": {
      "External id": 6445,"Record function id": 0, "finished": true, "Ev Idx": 6444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564858855.653, "dur": 9.694,
    "args": {
      "External id": 6446,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3289, "finished": true, "Ev Idx": 6445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564858869.664, "dur": 19.291,
    "args": {
      "External id": 6447,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3290, "finished": true, "Ev Idx": 6446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564858892.046, "dur": 49.092,
    "args": {
      "External id": 6448,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3291, "finished": true, "Ev Idx": 6447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564858943.713, "dur": 10.275,
    "args": {
      "External id": 6449,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3292, "finished": true, "Ev Idx": 6448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564858956.419, "dur": 7.624,
    "args": {
      "External id": 6450,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3293, "finished": true, "Ev Idx": 6449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564858967.250, "dur": 2.818,
    "args": {
      "External id": 6451,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3294, "finished": true, "Ev Idx": 6450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858968.591, "dur": 0.523,
    "args": {
      "External id": 6452,"Record function id": 0, "finished": true, "Ev Idx": 6451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564858971.783, "dur": 1.866,
    "args": {
      "External id": 6453,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3295, "finished": true, "Ev Idx": 6452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564858972.683, "dur": 0.383,
    "args": {
      "External id": 6454,"Record function id": 0, "finished": true, "Ev Idx": 6453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564858974.891, "dur": 9.488,
    "args": {
      "External id": 6455,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3296, "finished": true, "Ev Idx": 6454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564858986.610, "dur": 10.625,
    "args": {
      "External id": 6456,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3297, "finished": true, "Ev Idx": 6455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564858999.529, "dur": 8.358,
    "args": {
      "External id": 6457,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3298, "finished": true, "Ev Idx": 6456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564859009.405, "dur": 6.258,
    "args": {
      "External id": 6458,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3299, "finished": true, "Ev Idx": 6457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859040.364, "dur": 2.491,
    "args": {
      "External id": 6459,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3300, "finished": true, "Ev Idx": 6458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859041.575, "dur": 0.510,
    "args": {
      "External id": 6460,"Record function id": 0, "finished": true, "Ev Idx": 6459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859043.537, "dur": 1.638,
    "args": {
      "External id": 6461,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3301, "finished": true, "Ev Idx": 6460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859044.335, "dur": 0.342,
    "args": {
      "External id": 6462,"Record function id": 0, "finished": true, "Ev Idx": 6461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564859045.967, "dur": 2.699,
    "args": {
      "External id": 6463,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3302, "finished": true, "Ev Idx": 6462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859047.114, "dur": 0.846,
    "args": {
      "External id": 6464,"Record function id": 0, "finished": true, "Ev Idx": 6463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859049.353, "dur": 2.550,
    "args": {
      "External id": 6465,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3303, "finished": true, "Ev Idx": 6464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859050.977, "dur": 0.422,
    "args": {
      "External id": 6466,"Record function id": 0, "finished": true, "Ev Idx": 6465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859052.513, "dur": 2.215,
    "args": {
      "External id": 6467,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3304, "finished": true, "Ev Idx": 6466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859053.853, "dur": 0.426,
    "args": {
      "External id": 6468,"Record function id": 0, "finished": true, "Ev Idx": 6467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564859057.711, "dur": 3.256,
    "args": {
      "External id": 6469,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3305, "finished": true, "Ev Idx": 6468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859059.715, "dur": 0.413,
    "args": {
      "External id": 6470,"Record function id": 0, "finished": true, "Ev Idx": 6469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859063.515, "dur": 21.550,
    "args": {
      "External id": 6471,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3306, "finished": true, "Ev Idx": 6470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564859066.234, "dur": 16.472,
    "args": {
      "External id": 6472,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3306, "finished": true, "Ev Idx": 6471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564859068.145, "dur": 3.100,
    "args": {
      "External id": 6473,"Record function id": 0, "finished": true, "Ev Idx": 6472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859069.010, "dur": 1.881,
    "args": {
      "External id": 6474,"Record function id": 0, "finished": true, "Ev Idx": 6473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564859071.965, "dur": 10.162,
    "args": {
      "External id": 6475,"Record function id": 0, "finished": true, "Ev Idx": 6474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564859083.418, "dur": 1.383,
    "args": {
      "External id": 6476,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3307, "finished": true, "Ev Idx": 6475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859088.423, "dur": 2.742,
    "args": {
      "External id": 6477,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3308, "finished": true, "Ev Idx": 6476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859090.173, "dur": 0.423,
    "args": {
      "External id": 6478,"Record function id": 0, "finished": true, "Ev Idx": 6477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859091.825, "dur": 2.066,
    "args": {
      "External id": 6479,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3309, "finished": true, "Ev Idx": 6478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859093.124, "dur": 0.287,
    "args": {
      "External id": 6480,"Record function id": 0, "finished": true, "Ev Idx": 6479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564859094.404, "dur": 1.457,
    "args": {
      "External id": 6481,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3310, "finished": true, "Ev Idx": 6480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859095.095, "dur": 0.297,
    "args": {
      "External id": 6482,"Record function id": 0, "finished": true, "Ev Idx": 6481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859096.449, "dur": 1.475,
    "args": {
      "External id": 6483,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3311, "finished": true, "Ev Idx": 6482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859097.147, "dur": 0.303,
    "args": {
      "External id": 6484,"Record function id": 0, "finished": true, "Ev Idx": 6483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859098.520, "dur": 1.459,
    "args": {
      "External id": 6485,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3312, "finished": true, "Ev Idx": 6484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859099.203, "dur": 0.319,
    "args": {
      "External id": 6486,"Record function id": 0, "finished": true, "Ev Idx": 6485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564859101.608, "dur": 1.757,
    "args": {
      "External id": 6487,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3313, "finished": true, "Ev Idx": 6486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859102.603, "dur": 0.335,
    "args": {
      "External id": 6488,"Record function id": 0, "finished": true, "Ev Idx": 6487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859104.974, "dur": 13.486,
    "args": {
      "External id": 6489,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3314, "finished": true, "Ev Idx": 6488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564859105.906, "dur": 10.872,
    "args": {
      "External id": 6490,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3314, "finished": true, "Ev Idx": 6489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564859107.199, "dur": 2.376,
    "args": {
      "External id": 6491,"Record function id": 0, "finished": true, "Ev Idx": 6490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859107.865, "dur": 1.588,
    "args": {
      "External id": 6492,"Record function id": 0, "finished": true, "Ev Idx": 6491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564859109.881, "dur": 6.576,
    "args": {
      "External id": 6493,"Record function id": 0, "finished": true, "Ev Idx": 6492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564859117.161, "dur": 1.120,
    "args": {
      "External id": 6494,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3315, "finished": true, "Ev Idx": 6493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859120.743, "dur": 1.826,
    "args": {
      "External id": 6495,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3316, "finished": true, "Ev Idx": 6494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859121.654, "dur": 0.421,
    "args": {
      "External id": 6496,"Record function id": 0, "finished": true, "Ev Idx": 6495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564859124.233, "dur": 46.757,
    "args": {
      "External id": 6497,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3317, "finished": true, "Ev Idx": 6496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564859124.957, "dur": 1.625,
    "args": {
      "External id": 6498,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3317, "finished": true, "Ev Idx": 6497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859125.806, "dur": 0.342,
    "args": {
      "External id": 6499,"Record function id": 0, "finished": true, "Ev Idx": 6498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859126.910, "dur": 3.985,
    "args": {
      "External id": 6500,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3318, "finished": true, "Ev Idx": 6499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564859128.879, "dur": 1.872,
    "args": {
      "External id": 6501,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3318, "finished": true, "Ev Idx": 6500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564859131.368, "dur": 1.822,
    "args": {
      "External id": 6502,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3319, "finished": true, "Ev Idx": 6501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859132.567, "dur": 0.281,
    "args": {
      "External id": 6503,"Record function id": 0, "finished": true, "Ev Idx": 6502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859133.481, "dur": 2.513,
    "args": {
      "External id": 6504,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3320, "finished": true, "Ev Idx": 6503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564859134.522, "dur": 1.356,
    "args": {
      "External id": 6505,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3320, "finished": true, "Ev Idx": 6504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564859136.766, "dur": 32.196,
    "args": {
      "External id": 6506,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3321, "finished": true, "Ev Idx": 6505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859149.155, "dur": 1.067,
    "args": {
      "External id": 6507,"Record function id": 0, "finished": true, "Ev Idx": 6506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564859169.488, "dur": 1.147,
    "args": {
      "External id": 6508,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3322, "finished": true, "Ev Idx": 6507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564859174.312, "dur": 9.785,
    "args": {
      "External id": 6509,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3323, "finished": true, "Ev Idx": 6508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859187.142, "dur": 1.952,
    "args": {
      "External id": 6510,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3324, "finished": true, "Ev Idx": 6509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859187.927, "dur": 0.569,
    "args": {
      "External id": 6511,"Record function id": 0, "finished": true, "Ev Idx": 6510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859189.870, "dur": 1.406,
    "args": {
      "External id": 6512,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3324, "finished": true, "Ev Idx": 6511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859190.566, "dur": 0.300,
    "args": {
      "External id": 6513,"Record function id": 0, "finished": true, "Ev Idx": 6512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564859191.925, "dur": 0.982,
    "args": {
      "External id": 6514,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3324, "finished": true, "Ev Idx": 6513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859192.339, "dur": 0.245,
    "args": {
      "External id": 6515,"Record function id": 0, "finished": true, "Ev Idx": 6514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564859194.426, "dur": 7.436,
    "args": {
      "External id": 6516,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3324, "finished": true, "Ev Idx": 6515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564859206.960, "dur": 14.028,
    "args": {
      "External id": 6517,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3325, "finished": true, "Ev Idx": 6516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564859207.877, "dur": 0.408,
    "args": {
      "External id": 6518,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3325, "finished": true, "Ev Idx": 6517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564859209.061, "dur": 11.665,
    "args": {
      "External id": 6519,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3325, "finished": true, "Ev Idx": 6518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564859223.801, "dur": 0.177,
    "args": {
      "External id": 6520,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3326, "finished": true, "Ev Idx": 6519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564859234.740, "dur": 0.467,
    "args": {
      "External id": 6521,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3326, "finished": true, "Ev Idx": 6520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564859236.900, "dur": 41.068,
    "args": {
      "External id": 6522,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3326, "finished": true, "Ev Idx": 6521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564859237.646, "dur": 2.221,
    "args": {
      "External id": 6523,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3326, "finished": true, "Ev Idx": 6522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859238.940, "dur": 0.476,
    "args": {
      "External id": 6524,"Record function id": 0, "finished": true, "Ev Idx": 6523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859240.275, "dur": 2.324,
    "args": {
      "External id": 6525,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3327, "finished": true, "Ev Idx": 6524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859240.601, "dur": 1.895,
    "args": {
      "External id": 6526,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3327, "finished": true, "Ev Idx": 6525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564859243.279, "dur": 1.749,
    "args": {
      "External id": 6527,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3328, "finished": true, "Ev Idx": 6526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859244.056, "dur": 0.614,
    "args": {
      "External id": 6528,"Record function id": 0, "finished": true, "Ev Idx": 6527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859245.470, "dur": 1.385,
    "args": {
      "External id": 6529,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3329, "finished": true, "Ev Idx": 6528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859245.929, "dur": 0.851,
    "args": {
      "External id": 6530,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3329, "finished": true, "Ev Idx": 6529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564859247.390, "dur": 28.723,
    "args": {
      "External id": 6531,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3330, "finished": true, "Ev Idx": 6530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859258.818, "dur": 1.120,
    "args": {
      "External id": 6532,"Record function id": 0, "finished": true, "Ev Idx": 6531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564859276.648, "dur": 0.988,
    "args": {
      "External id": 6533,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3331, "finished": true, "Ev Idx": 6532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859279.973, "dur": 1.758,
    "args": {
      "External id": 6534,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3332, "finished": true, "Ev Idx": 6533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859280.898, "dur": 0.383,
    "args": {
      "External id": 6535,"Record function id": 0, "finished": true, "Ev Idx": 6534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564859284.348, "dur": 10.994,
    "args": {
      "External id": 6536,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3333, "finished": true, "Ev Idx": 6535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564859284.912, "dur": 10.289,
    "args": {
      "External id": 6537,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3333, "finished": true, "Ev Idx": 6536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564859285.929, "dur": 1.743,
    "args": {
      "External id": 6538,"Record function id": 0, "finished": true, "Ev Idx": 6537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859286.300, "dur": 1.216,
    "args": {
      "External id": 6539,"Record function id": 0, "finished": true, "Ev Idx": 6538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564859287.985, "dur": 6.782,
    "args": {
      "External id": 6540,"Record function id": 0, "finished": true, "Ev Idx": 6539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859300.849, "dur": 1.858,
    "args": {
      "External id": 6541,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3334, "finished": true, "Ev Idx": 6540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859301.363, "dur": 1.254,
    "args": {
      "External id": 6542,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3334, "finished": true, "Ev Idx": 6541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564859316.033, "dur": 36.555,
    "args": {
      "External id": 6543,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3335, "finished": true, "Ev Idx": 6542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564859316.746, "dur": 3.015,
    "args": {
      "External id": 6544,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3335, "finished": true, "Ev Idx": 6543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859318.327, "dur": 0.849,
    "args": {
      "External id": 6545,"Record function id": 0, "finished": true, "Ev Idx": 6544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859318.727, "dur": 0.341,
    "args": {
      "External id": 6546,"Record function id": 0, "finished": true, "Ev Idx": 6545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564859320.288, "dur": 32.158,
    "args": {
      "External id": 6547,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3336, "finished": true, "Ev Idx": 6546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859320.731, "dur": 1.783,
    "args": {
      "External id": 6548,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3336, "finished": true, "Ev Idx": 6547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859321.103, "dur": 1.325,
    "args": {
      "External id": 6549,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3336, "finished": true, "Ev Idx": 6548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564859323.021, "dur": 27.389,
    "args": {
      "External id": 6550,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3337, "finished": true, "Ev Idx": 6549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859334.995, "dur": 1.232,
    "args": {
      "External id": 6551,"Record function id": 0, "finished": true, "Ev Idx": 6550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564859350.983, "dur": 1.144,
    "args": {
      "External id": 6552,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3338, "finished": true, "Ev Idx": 6551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564859357.305, "dur": 9.811,
    "args": {
      "External id": 6553,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3339, "finished": true, "Ev Idx": 6552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564859375.937, "dur": 0.517,
    "args": {
      "External id": 6554,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3340, "finished": true, "Ev Idx": 6553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564859378.584, "dur": 11.742,
    "args": {
      "External id": 6555,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3340, "finished": true, "Ev Idx": 6554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564859380.770, "dur": 0.564,
    "args": {
      "External id": 6556,"Record function id": 0, "finished": true, "Ev Idx": 6555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564859381.921, "dur": 0.067,
    "args": {
      "External id": 6557,"Record function id": 0, "finished": true, "Ev Idx": 6556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564859392.983, "dur": 14.743,
    "args": {
      "External id": 6558,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3341, "finished": true, "Ev Idx": 6557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564859411.244, "dur": 7.668,
    "args": {
      "External id": 6559,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3342, "finished": true, "Ev Idx": 6558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564859420.931, "dur": 11.850,
    "args": {
      "External id": 6560,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3343, "finished": true, "Ev Idx": 6559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564859434.458, "dur": 8.808,
    "args": {
      "External id": 6561,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3344, "finished": true, "Ev Idx": 6560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564859445.671, "dur": 0.123,
    "args": {
      "External id": 6562,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3345, "finished": true, "Ev Idx": 6561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564859446.646, "dur": 7.531,
    "args": {
      "External id": 6563,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3345, "finished": true, "Ev Idx": 6562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564859467.873, "dur": 44.704,
    "args": {
      "External id": 6564,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3346, "finished": true, "Ev Idx": 6563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564859468.571, "dur": 3.124,
    "args": {
      "External id": 6565,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3346, "finished": true, "Ev Idx": 6564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859469.660, "dur": 1.414,
    "args": {
      "External id": 6566,"Record function id": 0, "finished": true, "Ev Idx": 6565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859470.354, "dur": 0.595,
    "args": {
      "External id": 6567,"Record function id": 0, "finished": true, "Ev Idx": 6566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564859472.221, "dur": 40.229,
    "args": {
      "External id": 6568,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3347, "finished": true, "Ev Idx": 6567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859472.691, "dur": 2.563,
    "args": {
      "External id": 6569,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3347, "finished": true, "Ev Idx": 6568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859473.097, "dur": 2.067,
    "args": {
      "External id": 6570,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3347, "finished": true, "Ev Idx": 6569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564859475.573, "dur": 35.027,
    "args": {
      "External id": 6571,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3348, "finished": true, "Ev Idx": 6570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859485.934, "dur": 1.120,
    "args": {
      "External id": 6572,"Record function id": 0, "finished": true, "Ev Idx": 6571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564859511.205, "dur": 0.969,
    "args": {
      "External id": 6573,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3349, "finished": true, "Ev Idx": 6572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564859521.010, "dur": 61.354,
    "args": {
      "External id": 6574,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3350, "finished": true, "Ev Idx": 6573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564859590.909, "dur": 43.774,
    "args": {
      "External id": 6575,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3351, "finished": true, "Ev Idx": 6574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564859591.302, "dur": 3.167,
    "args": {
      "External id": 6576,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3351, "finished": true, "Ev Idx": 6575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859592.546, "dur": 1.323,
    "args": {
      "External id": 6577,"Record function id": 0, "finished": true, "Ev Idx": 6576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859593.206, "dur": 0.521,
    "args": {
      "External id": 6578,"Record function id": 0, "finished": true, "Ev Idx": 6577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564859594.907, "dur": 39.645,
    "args": {
      "External id": 6579,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3352, "finished": true, "Ev Idx": 6578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859595.419, "dur": 2.019,
    "args": {
      "External id": 6580,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3352, "finished": true, "Ev Idx": 6579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859595.722, "dur": 1.618,
    "args": {
      "External id": 6581,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3352, "finished": true, "Ev Idx": 6580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564859597.740, "dur": 34.858,
    "args": {
      "External id": 6582,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3353, "finished": true, "Ev Idx": 6581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859608.947, "dur": 1.081,
    "args": {
      "External id": 6583,"Record function id": 0, "finished": true, "Ev Idx": 6582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564859633.251, "dur": 0.994,
    "args": {
      "External id": 6584,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3354, "finished": true, "Ev Idx": 6583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564859636.671, "dur": 46.024,
    "args": {
      "External id": 6585,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3355, "finished": true, "Ev Idx": 6584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564859690.755, "dur": 44.983,
    "args": {
      "External id": 6586,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3356, "finished": true, "Ev Idx": 6585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564859691.320, "dur": 4.161,
    "args": {
      "External id": 6587,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3356, "finished": true, "Ev Idx": 6586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859693.354, "dur": 1.283,
    "args": {
      "External id": 6588,"Record function id": 0, "finished": true, "Ev Idx": 6587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859694.036, "dur": 0.472,
    "args": {
      "External id": 6589,"Record function id": 0, "finished": true, "Ev Idx": 6588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564859695.830, "dur": 39.809,
    "args": {
      "External id": 6590,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3357, "finished": true, "Ev Idx": 6589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859696.281, "dur": 1.568,
    "args": {
      "External id": 6591,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3357, "finished": true, "Ev Idx": 6590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859696.608, "dur": 1.127,
    "args": {
      "External id": 6592,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3357, "finished": true, "Ev Idx": 6591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564859698.171, "dur": 35.483,
    "args": {
      "External id": 6593,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3358, "finished": true, "Ev Idx": 6592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859709.253, "dur": 1.118,
    "args": {
      "External id": 6594,"Record function id": 0, "finished": true, "Ev Idx": 6593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564859734.242, "dur": 1.061,
    "args": {
      "External id": 6595,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3359, "finished": true, "Ev Idx": 6594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564859738.739, "dur": 8.752,
    "args": {
      "External id": 6596,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3360, "finished": true, "Ev Idx": 6595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564859768.715, "dur": 0.376,
    "args": {
      "External id": 6597,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3361, "finished": true, "Ev Idx": 6596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564859771.124, "dur": 10.136,
    "args": {
      "External id": 6598,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3361, "finished": true, "Ev Idx": 6597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564859772.547, "dur": 0.443,
    "args": {
      "External id": 6599,"Record function id": 0, "finished": true, "Ev Idx": 6598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564859773.279, "dur": 0.249,
    "args": {
      "External id": 6600,"Record function id": 0, "finished": true, "Ev Idx": 6599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564859783.838, "dur": 13.360,
    "args": {
      "External id": 6601,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3362, "finished": true, "Ev Idx": 6600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564859800.799, "dur": 8.318,
    "args": {
      "External id": 6602,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3363, "finished": true, "Ev Idx": 6601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564859811.004, "dur": 8.039,
    "args": {
      "External id": 6603,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3364, "finished": true, "Ev Idx": 6602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564859820.727, "dur": 8.902,
    "args": {
      "External id": 6604,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3365, "finished": true, "Ev Idx": 6603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564859832.122, "dur": 0.095,
    "args": {
      "External id": 6605,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3366, "finished": true, "Ev Idx": 6604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564859833.073, "dur": 7.946,
    "args": {
      "External id": 6606,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3366, "finished": true, "Ev Idx": 6605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564859860.162, "dur": 38.596,
    "args": {
      "External id": 6607,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3367, "finished": true, "Ev Idx": 6606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564859860.765, "dur": 3.787,
    "args": {
      "External id": 6608,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3367, "finished": true, "Ev Idx": 6607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859862.546, "dur": 1.345,
    "args": {
      "External id": 6609,"Record function id": 0, "finished": true, "Ev Idx": 6608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859863.251, "dur": 0.494,
    "args": {
      "External id": 6610,"Record function id": 0, "finished": true, "Ev Idx": 6609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564859864.913, "dur": 33.724,
    "args": {
      "External id": 6611,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3368, "finished": true, "Ev Idx": 6610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859865.445, "dur": 1.698,
    "args": {
      "External id": 6612,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3368, "finished": true, "Ev Idx": 6611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859865.907, "dur": 1.115,
    "args": {
      "External id": 6613,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3368, "finished": true, "Ev Idx": 6612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564859867.473, "dur": 28.976,
    "args": {
      "External id": 6614,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3369, "finished": true, "Ev Idx": 6613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859879.645, "dur": 1.307,
    "args": {
      "External id": 6615,"Record function id": 0, "finished": true, "Ev Idx": 6614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564859897.057, "dur": 1.247,
    "args": {
      "External id": 6616,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3370, "finished": true, "Ev Idx": 6615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859902.050, "dur": 1.331,
    "args": {
      "External id": 6617,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3371, "finished": true, "Ev Idx": 6616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859905.344, "dur": 2.769,
    "args": {
      "External id": 6618,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3372, "finished": true, "Ev Idx": 6617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859906.514, "dur": 0.666,
    "args": {
      "External id": 6619,"Record function id": 0, "finished": true, "Ev Idx": 6618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564859915.207, "dur": 40.737,
    "args": {
      "External id": 6620,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3373, "finished": true, "Ev Idx": 6619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564859915.798, "dur": 2.283,
    "args": {
      "External id": 6621,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3373, "finished": true, "Ev Idx": 6620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859916.733, "dur": 0.889,
    "args": {
      "External id": 6622,"Record function id": 0, "finished": true, "Ev Idx": 6621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859917.240, "dur": 0.288,
    "args": {
      "External id": 6623,"Record function id": 0, "finished": true, "Ev Idx": 6622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564859918.418, "dur": 37.430,
    "args": {
      "External id": 6624,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3374, "finished": true, "Ev Idx": 6623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859918.992, "dur": 1.278,
    "args": {
      "External id": 6625,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3374, "finished": true, "Ev Idx": 6624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859919.269, "dur": 0.891,
    "args": {
      "External id": 6626,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3374, "finished": true, "Ev Idx": 6625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564859920.573, "dur": 33.340,
    "args": {
      "External id": 6627,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3375, "finished": true, "Ev Idx": 6626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859931.813, "dur": 1.113,
    "args": {
      "External id": 6628,"Record function id": 0, "finished": true, "Ev Idx": 6627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564859954.582, "dur": 0.973,
    "args": {
      "External id": 6629,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3376, "finished": true, "Ev Idx": 6628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859958.420, "dur": 1.345,
    "args": {
      "External id": 6630,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3377, "finished": true, "Ev Idx": 6629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859961.275, "dur": 1.918,
    "args": {
      "External id": 6631,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3378, "finished": true, "Ev Idx": 6630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859962.081, "dur": 0.605,
    "args": {
      "External id": 6632,"Record function id": 0, "finished": true, "Ev Idx": 6631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564859969.698, "dur": 36.701,
    "args": {
      "External id": 6633,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3379, "finished": true, "Ev Idx": 6632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564859970.135, "dur": 2.469,
    "args": {
      "External id": 6634,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3379, "finished": true, "Ev Idx": 6633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564859971.247, "dur": 0.956,
    "args": {
      "External id": 6635,"Record function id": 0, "finished": true, "Ev Idx": 6634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564859971.784, "dur": 0.308,
    "args": {
      "External id": 6636,"Record function id": 0, "finished": true, "Ev Idx": 6635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564859972.897, "dur": 33.428,
    "args": {
      "External id": 6637,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3380, "finished": true, "Ev Idx": 6636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564859973.641, "dur": 1.371,
    "args": {
      "External id": 6638,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3380, "finished": true, "Ev Idx": 6637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564859973.908, "dur": 1.008,
    "args": {
      "External id": 6639,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3380, "finished": true, "Ev Idx": 6638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564859975.316, "dur": 29.113,
    "args": {
      "External id": 6640,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3381, "finished": true, "Ev Idx": 6639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564859985.582, "dur": 1.038,
    "args": {
      "External id": 6641,"Record function id": 0, "finished": true, "Ev Idx": 6640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564860004.952, "dur": 1.038,
    "args": {
      "External id": 6642,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3382, "finished": true, "Ev Idx": 6641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564860008.923, "dur": 1.145,
    "args": {
      "External id": 6643,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3383, "finished": true, "Ev Idx": 6642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564860011.433, "dur": 1.792,
    "args": {
      "External id": 6644,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3384, "finished": true, "Ev Idx": 6643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860012.184, "dur": 0.455,
    "args": {
      "External id": 6645,"Record function id": 0, "finished": true, "Ev Idx": 6644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564860016.111, "dur": 2.143,
    "args": {
      "External id": 6646,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3385, "finished": true, "Ev Idx": 6645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860017.344, "dur": 0.361,
    "args": {
      "External id": 6647,"Record function id": 0, "finished": true, "Ev Idx": 6646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564860019.423, "dur": 1.056,
    "args": {
      "External id": 6648,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3385, "finished": true, "Ev Idx": 6647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860019.861, "dur": 0.280,
    "args": {
      "External id": 6649,"Record function id": 0, "finished": true, "Ev Idx": 6648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564860021.307, "dur": 11.864,
    "args": {
      "External id": 6650,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3385, "finished": true, "Ev Idx": 6649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860038.682, "dur": 4.964,
    "args": {
      "External id": 6651,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3386, "finished": true, "Ev Idx": 6650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860041.575, "dur": 0.657,
    "args": {
      "External id": 6652,"Record function id": 0, "finished": true, "Ev Idx": 6651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860045.937, "dur": 2.018,
    "args": {
      "External id": 6653,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3387, "finished": true, "Ev Idx": 6652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860047.025, "dur": 0.351,
    "args": {
      "External id": 6654,"Record function id": 0, "finished": true, "Ev Idx": 6653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564860049.947, "dur": 10.550,
    "args": {
      "External id": 6655,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3388, "finished": true, "Ev Idx": 6654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564860064.562, "dur": 16.890,
    "args": {
      "External id": 6656,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3389, "finished": true, "Ev Idx": 6655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564860109.433, "dur": 10.112,
    "args": {
      "External id": 6657,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3390, "finished": true, "Ev Idx": 6656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564860121.343, "dur": 9.621,
    "args": {
      "External id": 6658,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3391, "finished": true, "Ev Idx": 6657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564860133.060, "dur": 7.760,
    "args": {
      "External id": 6659,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3392, "finished": true, "Ev Idx": 6658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860143.437, "dur": 2.337,
    "args": {
      "External id": 6660,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3393, "finished": true, "Ev Idx": 6659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860144.541, "dur": 0.507,
    "args": {
      "External id": 6661,"Record function id": 0, "finished": true, "Ev Idx": 6660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860147.261, "dur": 1.691,
    "args": {
      "External id": 6662,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3394, "finished": true, "Ev Idx": 6661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860148.102, "dur": 0.341,
    "args": {
      "External id": 6663,"Record function id": 0, "finished": true, "Ev Idx": 6662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 537, "tid": 537,
    "ts": 7429564860149.924, "dur": 6.532,
    "args": {
      "External id": 6664,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3395, "finished": true, "Ev Idx": 6663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 537, "tid": 537,
    "ts": 7429564860158.168, "dur": 9.116,
    "args": {
      "External id": 6665,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3396, "finished": true, "Ev Idx": 6664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564860169.490, "dur": 7.642,
    "args": {
      "External id": 6666,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3397, "finished": true, "Ev Idx": 6665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564860178.454, "dur": 6.866,
    "args": {
      "External id": 6667,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3398, "finished": true, "Ev Idx": 6666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860208.817, "dur": 2.483,
    "args": {
      "External id": 6668,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3399, "finished": true, "Ev Idx": 6667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860210.107, "dur": 0.517,
    "args": {
      "External id": 6669,"Record function id": 0, "finished": true, "Ev Idx": 6668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860211.962, "dur": 1.572,
    "args": {
      "External id": 6670,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3400, "finished": true, "Ev Idx": 6669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860212.740, "dur": 0.308,
    "args": {
      "External id": 6671,"Record function id": 0, "finished": true, "Ev Idx": 6670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564860214.309, "dur": 2.232,
    "args": {
      "External id": 6672,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3401, "finished": true, "Ev Idx": 6671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860215.263, "dur": 0.607,
    "args": {
      "External id": 6673,"Record function id": 0, "finished": true, "Ev Idx": 6672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860217.190, "dur": 2.855,
    "args": {
      "External id": 6674,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3402, "finished": true, "Ev Idx": 6673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860219.155, "dur": 0.390,
    "args": {
      "External id": 6675,"Record function id": 0, "finished": true, "Ev Idx": 6674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860220.620, "dur": 2.079,
    "args": {
      "External id": 6676,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3403, "finished": true, "Ev Idx": 6675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860221.942, "dur": 0.296,
    "args": {
      "External id": 6677,"Record function id": 0, "finished": true, "Ev Idx": 6676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564860225.638, "dur": 2.987,
    "args": {
      "External id": 6678,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3404, "finished": true, "Ev Idx": 6677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860227.494, "dur": 0.419,
    "args": {
      "External id": 6679,"Record function id": 0, "finished": true, "Ev Idx": 6678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860230.972, "dur": 19.441,
    "args": {
      "External id": 6680,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3405, "finished": true, "Ev Idx": 6679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564860233.572, "dur": 14.778,
    "args": {
      "External id": 6681,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3405, "finished": true, "Ev Idx": 6680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564860235.145, "dur": 2.675,
    "args": {
      "External id": 6682,"Record function id": 0, "finished": true, "Ev Idx": 6681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564860235.993, "dur": 1.617,
    "args": {
      "External id": 6683,"Record function id": 0, "finished": true, "Ev Idx": 6682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564860238.465, "dur": 9.292,
    "args": {
      "External id": 6684,"Record function id": 0, "finished": true, "Ev Idx": 6683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564860248.858, "dur": 1.277,
    "args": {
      "External id": 6685,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3406, "finished": true, "Ev Idx": 6684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860253.646, "dur": 2.400,
    "args": {
      "External id": 6686,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3407, "finished": true, "Ev Idx": 6685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860254.967, "dur": 0.487,
    "args": {
      "External id": 6687,"Record function id": 0, "finished": true, "Ev Idx": 6686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860256.665, "dur": 2.140,
    "args": {
      "External id": 6688,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3408, "finished": true, "Ev Idx": 6687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860258.023, "dur": 0.290,
    "args": {
      "External id": 6689,"Record function id": 0, "finished": true, "Ev Idx": 6688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 537, "tid": 537,
    "ts": 7429564860259.347, "dur": 2.120,
    "args": {
      "External id": 6690,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3409, "finished": true, "Ev Idx": 6689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860260.665, "dur": 0.306,
    "args": {
      "External id": 6691,"Record function id": 0, "finished": true, "Ev Idx": 6690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860262.076, "dur": 1.530,
    "args": {
      "External id": 6692,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3410, "finished": true, "Ev Idx": 6691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860262.794, "dur": 0.312,
    "args": {
      "External id": 6693,"Record function id": 0, "finished": true, "Ev Idx": 6692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860264.218, "dur": 1.407,
    "args": {
      "External id": 6694,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3411, "finished": true, "Ev Idx": 6693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860264.885, "dur": 0.264,
    "args": {
      "External id": 6695,"Record function id": 0, "finished": true, "Ev Idx": 6694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564860267.130, "dur": 1.549,
    "args": {
      "External id": 6696,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3412, "finished": true, "Ev Idx": 6695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860267.909, "dur": 0.340,
    "args": {
      "External id": 6697,"Record function id": 0, "finished": true, "Ev Idx": 6696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860270.097, "dur": 12.740,
    "args": {
      "External id": 6698,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3413, "finished": true, "Ev Idx": 6697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564860270.993, "dur": 9.875,
    "args": {
      "External id": 6699,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3413, "finished": true, "Ev Idx": 6698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564860272.175, "dur": 1.813,
    "args": {
      "External id": 6700,"Record function id": 0, "finished": true, "Ev Idx": 6699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564860272.469, "dur": 1.382,
    "args": {
      "External id": 6701,"Record function id": 0, "finished": true, "Ev Idx": 6700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564860274.401, "dur": 6.146,
    "args": {
      "External id": 6702,"Record function id": 0, "finished": true, "Ev Idx": 6701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564860281.263, "dur": 1.402,
    "args": {
      "External id": 6703,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3414, "finished": true, "Ev Idx": 6702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564860284.984, "dur": 1.733,
    "args": {
      "External id": 6704,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3415, "finished": true, "Ev Idx": 6703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860285.815, "dur": 0.412,
    "args": {
      "External id": 6705,"Record function id": 0, "finished": true, "Ev Idx": 6704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564860288.386, "dur": 46.005,
    "args": {
      "External id": 6706,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3416, "finished": true, "Ev Idx": 6705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564860288.995, "dur": 1.545,
    "args": {
      "External id": 6707,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3416, "finished": true, "Ev Idx": 6706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860289.813, "dur": 0.312,
    "args": {
      "External id": 6708,"Record function id": 0, "finished": true, "Ev Idx": 6707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860290.878, "dur": 3.938,
    "args": {
      "External id": 6709,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3417, "finished": true, "Ev Idx": 6708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564860292.721, "dur": 1.940,
    "args": {
      "External id": 6710,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3417, "finished": true, "Ev Idx": 6709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564860295.306, "dur": 1.308,
    "args": {
      "External id": 6711,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3418, "finished": true, "Ev Idx": 6710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860295.980, "dur": 0.274,
    "args": {
      "External id": 6712,"Record function id": 0, "finished": true, "Ev Idx": 6711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860296.916, "dur": 3.255,
    "args": {
      "External id": 6713,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3419, "finished": true, "Ev Idx": 6712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 537, "tid": 537,
    "ts": 7429564860297.765, "dur": 2.311,
    "args": {
      "External id": 6714,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3419, "finished": true, "Ev Idx": 6713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564860300.907, "dur": 31.274,
    "args": {
      "External id": 6715,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3420, "finished": true, "Ev Idx": 6714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564860315.454, "dur": 1.052,
    "args": {
      "External id": 6716,"Record function id": 0, "finished": true, "Ev Idx": 6715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564860332.826, "dur": 1.180,
    "args": {
      "External id": 6717,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3421, "finished": true, "Ev Idx": 6716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564860337.633, "dur": 10.193,
    "args": {
      "External id": 6718,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3422, "finished": true, "Ev Idx": 6717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860350.723, "dur": 1.809,
    "args": {
      "External id": 6719,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3423, "finished": true, "Ev Idx": 6718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860351.449, "dur": 0.406,
    "args": {
      "External id": 6720,"Record function id": 0, "finished": true, "Ev Idx": 6719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860353.213, "dur": 1.372,
    "args": {
      "External id": 6721,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3423, "finished": true, "Ev Idx": 6720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860353.860, "dur": 0.294,
    "args": {
      "External id": 6722,"Record function id": 0, "finished": true, "Ev Idx": 6721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564860355.256, "dur": 0.998,
    "args": {
      "External id": 6723,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3423, "finished": true, "Ev Idx": 6722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860355.709, "dur": 0.233,
    "args": {
      "External id": 6724,"Record function id": 0, "finished": true, "Ev Idx": 6723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564860357.838, "dur": 8.580,
    "args": {
      "External id": 6725,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3423, "finished": true, "Ev Idx": 6724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::softmax", "pid": 537, "tid": 537,
    "ts": 7429564860371.903, "dur": 13.112,
    "args": {
      "External id": 6726,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3424, "finished": true, "Ev Idx": 6725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564860372.957, "dur": 0.381,
    "args": {
      "External id": 6727,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3424, "finished": true, "Ev Idx": 6726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_softmax", "pid": 537, "tid": 537,
    "ts": 7429564860373.997, "dur": 10.744,
    "args": {
      "External id": 6728,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3424, "finished": true, "Ev Idx": 6727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564860387.781, "dur": 0.202,
    "args": {
      "External id": 6729,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3425, "finished": true, "Ev Idx": 6728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::dropout", "pid": 537, "tid": 537,
    "ts": 7429564860398.636, "dur": 0.482,
    "args": {
      "External id": 6730,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3425, "finished": true, "Ev Idx": 6729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564860400.858, "dur": 40.486,
    "args": {
      "External id": 6731,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3425, "finished": true, "Ev Idx": 6730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564860401.379, "dur": 2.590,
    "args": {
      "External id": 6732,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3425, "finished": true, "Ev Idx": 6731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860403.040, "dur": 0.412,
    "args": {
      "External id": 6733,"Record function id": 0, "finished": true, "Ev Idx": 6732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860404.383, "dur": 2.429,
    "args": {
      "External id": 6734,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3426, "finished": true, "Ev Idx": 6733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564860404.704, "dur": 2.005,
    "args": {
      "External id": 6735,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3426, "finished": true, "Ev Idx": 6734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 537, "tid": 537,
    "ts": 7429564860407.221, "dur": 1.260,
    "args": {
      "External id": 6736,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3427, "finished": true, "Ev Idx": 6735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860407.828, "dur": 0.282,
    "args": {
      "External id": 6737,"Record function id": 0, "finished": true, "Ev Idx": 6736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860408.765, "dur": 1.127,
    "args": {
      "External id": 6738,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3428, "finished": true, "Ev Idx": 6737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564860409.015, "dur": 0.805,
    "args": {
      "External id": 6739,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3428, "finished": true, "Ev Idx": 6738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 537, "tid": 537,
    "ts": 7429564860410.292, "dur": 28.990,
    "args": {
      "External id": 6740,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3429, "finished": true, "Ev Idx": 6739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564860422.237, "dur": 1.132,
    "args": {
      "External id": 6741,"Record function id": 0, "finished": true, "Ev Idx": 6740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564860439.897, "dur": 1.130,
    "args": {
      "External id": 6742,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3430, "finished": true, "Ev Idx": 6741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564860443.359, "dur": 1.756,
    "args": {
      "External id": 6743,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3431, "finished": true, "Ev Idx": 6742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860444.226, "dur": 0.383,
    "args": {
      "External id": 6744,"Record function id": 0, "finished": true, "Ev Idx": 6743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 537, "tid": 537,
    "ts": 7429564860447.747, "dur": 11.238,
    "args": {
      "External id": 6745,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3432, "finished": true, "Ev Idx": 6744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 537, "tid": 537,
    "ts": 7429564860448.271, "dur": 10.574,
    "args": {
      "External id": 6746,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3432, "finished": true, "Ev Idx": 6745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 537, "tid": 537,
    "ts": 7429564860449.342, "dur": 2.032,
    "args": {
      "External id": 6747,"Record function id": 0, "finished": true, "Ev Idx": 6746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564860449.748, "dur": 1.474,
    "args": {
      "External id": 6748,"Record function id": 0, "finished": true, "Ev Idx": 6747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 537, "tid": 537,
    "ts": 7429564860451.708, "dur": 6.694,
    "args": {
      "External id": 6749,"Record function id": 0, "finished": true, "Ev Idx": 6748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860464.866, "dur": 1.642,
    "args": {
      "External id": 6750,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3433, "finished": true, "Ev Idx": 6749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564860465.176, "dur": 1.239,
    "args": {
      "External id": 6751,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3433, "finished": true, "Ev Idx": 6750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564860479.034, "dur": 34.456,
    "args": {
      "External id": 6752,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3434, "finished": true, "Ev Idx": 6751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564860479.592, "dur": 2.854,
    "args": {
      "External id": 6753,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3434, "finished": true, "Ev Idx": 6752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564860480.979, "dur": 0.962,
    "args": {
      "External id": 6754,"Record function id": 0, "finished": true, "Ev Idx": 6753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860481.482, "dur": 0.361,
    "args": {
      "External id": 6755,"Record function id": 0, "finished": true, "Ev Idx": 6754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564860482.816, "dur": 30.534,
    "args": {
      "External id": 6756,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3435, "finished": true, "Ev Idx": 6755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860483.214, "dur": 1.554,
    "args": {
      "External id": 6757,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3435, "finished": true, "Ev Idx": 6756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564860483.608, "dur": 1.072,
    "args": {
      "External id": 6758,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3435, "finished": true, "Ev Idx": 6757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564860485.204, "dur": 26.235,
    "args": {
      "External id": 6759,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3436, "finished": true, "Ev Idx": 6758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564860497.336, "dur": 1.026,
    "args": {
      "External id": 6760,"Record function id": 0, "finished": true, "Ev Idx": 6759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564860512.030, "dur": 0.976,
    "args": {
      "External id": 6761,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3437, "finished": true, "Ev Idx": 6760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564860518.126, "dur": 8.406,
    "args": {
      "External id": 6762,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3438, "finished": true, "Ev Idx": 6761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564860535.276, "dur": 0.199,
    "args": {
      "External id": 6763,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3439, "finished": true, "Ev Idx": 6762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564860537.501, "dur": 11.081,
    "args": {
      "External id": 6764,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3439, "finished": true, "Ev Idx": 6763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564860539.200, "dur": 0.395,
    "args": {
      "External id": 6765,"Record function id": 0, "finished": true, "Ev Idx": 6764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564860539.962, "dur": 0.070,
    "args": {
      "External id": 6766,"Record function id": 0, "finished": true, "Ev Idx": 6765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564860551.194, "dur": 11.318,
    "args": {
      "External id": 6767,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3440, "finished": true, "Ev Idx": 6766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564860565.974, "dur": 8.799,
    "args": {
      "External id": 6768,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3441, "finished": true, "Ev Idx": 6767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564860576.641, "dur": 8.518,
    "args": {
      "External id": 6769,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3442, "finished": true, "Ev Idx": 6768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564860586.834, "dur": 7.994,
    "args": {
      "External id": 6770,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3443, "finished": true, "Ev Idx": 6769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564860597.334, "dur": 0.102,
    "args": {
      "External id": 6771,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3444, "finished": true, "Ev Idx": 6770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564860598.376, "dur": 8.944,
    "args": {
      "External id": 6772,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3444, "finished": true, "Ev Idx": 6771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564860620.540, "dur": 114.747,
    "args": {
      "External id": 6773,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3445, "finished": true, "Ev Idx": 6772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564860620.969, "dur": 4.144,
    "args": {
      "External id": 6774,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3445, "finished": true, "Ev Idx": 6773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564860623.499, "dur": 1.002,
    "args": {
      "External id": 6775,"Record function id": 0, "finished": true, "Ev Idx": 6774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860623.946, "dur": 0.432,
    "args": {
      "External id": 6776,"Record function id": 0, "finished": true, "Ev Idx": 6775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564860625.429, "dur": 109.718,
    "args": {
      "External id": 6777,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3446, "finished": true, "Ev Idx": 6776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860625.807, "dur": 2.285,
    "args": {
      "External id": 6778,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3446, "finished": true, "Ev Idx": 6777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564860626.095, "dur": 1.905,
    "args": {
      "External id": 6779,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3446, "finished": true, "Ev Idx": 6778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564860628.543, "dur": 104.184,
    "args": {
      "External id": 6780,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3447, "finished": true, "Ev Idx": 6779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564860703.303, "dur": 1.332,
    "args": {
      "External id": 6781,"Record function id": 0, "finished": true, "Ev Idx": 6780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564860733.519, "dur": 1.213,
    "args": {
      "External id": 6782,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3448, "finished": true, "Ev Idx": 6781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::silu", "pid": 537, "tid": 537,
    "ts": 7429564860744.560, "dur": 10.209,
    "args": {
      "External id": 6783,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3449, "finished": true, "Ev Idx": 6782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564860762.924, "dur": 78.386,
    "args": {
      "External id": 6784,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3450, "finished": true, "Ev Idx": 6783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564860763.337, "dur": 3.012,
    "args": {
      "External id": 6785,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3450, "finished": true, "Ev Idx": 6784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564860764.539, "dur": 1.246,
    "args": {
      "External id": 6786,"Record function id": 0, "finished": true, "Ev Idx": 6785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860765.146, "dur": 0.439,
    "args": {
      "External id": 6787,"Record function id": 0, "finished": true, "Ev Idx": 6786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564860766.684, "dur": 74.522,
    "args": {
      "External id": 6788,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3451, "finished": true, "Ev Idx": 6787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860767.090, "dur": 3.215,
    "args": {
      "External id": 6789,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3451, "finished": true, "Ev Idx": 6788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564860767.593, "dur": 2.619,
    "args": {
      "External id": 6790,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3451, "finished": true, "Ev Idx": 6789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564860770.638, "dur": 68.524,
    "args": {
      "External id": 6791,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3452, "finished": true, "Ev Idx": 6790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564860815.867, "dur": 1.138,
    "args": {
      "External id": 6792,"Record function id": 0, "finished": true, "Ev Idx": 6791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564860839.716, "dur": 1.133,
    "args": {
      "External id": 6793,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3453, "finished": true, "Ev Idx": 6792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564860843.707, "dur": 9.187,
    "args": {
      "External id": 6794,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3454, "finished": true, "Ev Idx": 6793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564860859.895, "dur": 44.644,
    "args": {
      "External id": 6795,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3455, "finished": true, "Ev Idx": 6794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564860860.389, "dur": 2.636,
    "args": {
      "External id": 6796,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3455, "finished": true, "Ev Idx": 6795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564860861.328, "dur": 1.079,
    "args": {
      "External id": 6797,"Record function id": 0, "finished": true, "Ev Idx": 6796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564860861.748, "dur": 0.506,
    "args": {
      "External id": 6798,"Record function id": 0, "finished": true, "Ev Idx": 6797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564860863.469, "dur": 40.970,
    "args": {
      "External id": 6799,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3456, "finished": true, "Ev Idx": 6798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564860863.973, "dur": 1.655,
    "args": {
      "External id": 6800,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3456, "finished": true, "Ev Idx": 6799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564860864.425, "dur": 1.090,
    "args": {
      "External id": 6801,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3456, "finished": true, "Ev Idx": 6800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564860866.021, "dur": 36.567,
    "args": {
      "External id": 6802,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3457, "finished": true, "Ev Idx": 6801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564860880.641, "dur": 1.100,
    "args": {
      "External id": 6803,"Record function id": 0, "finished": true, "Ev Idx": 6802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564860903.152, "dur": 0.996,
    "args": {
      "External id": 6804,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3458, "finished": true, "Ev Idx": 6803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564860907.244, "dur": 8.500,
    "args": {
      "External id": 6805,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3459, "finished": true, "Ev Idx": 6804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564860931.880, "dur": 0.188,
    "args": {
      "External id": 6806,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3460, "finished": true, "Ev Idx": 6805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 537, "tid": 537,
    "ts": 7429564860934.152, "dur": 9.987,
    "args": {
      "External id": 6807,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3460, "finished": true, "Ev Idx": 6806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 537, "tid": 537,
    "ts": 7429564860935.495, "dur": 0.485,
    "args": {
      "External id": 6808,"Record function id": 0, "finished": true, "Ev Idx": 6807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564860936.377, "dur": 0.098,
    "args": {
      "External id": 6809,"Record function id": 0, "finished": true, "Ev Idx": 6808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 537, "tid": 537,
    "ts": 7429564860946.458, "dur": 13.002,
    "args": {
      "External id": 6810,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3461, "finished": true, "Ev Idx": 6809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 537, "tid": 537,
    "ts": 7429564860962.939, "dur": 7.713,
    "args": {
      "External id": 6811,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3462, "finished": true, "Ev Idx": 6810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::rsqrt", "pid": 537, "tid": 537,
    "ts": 7429564860972.369, "dur": 10.451,
    "args": {
      "External id": 6812,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3463, "finished": true, "Ev Idx": 6811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564860984.405, "dur": 7.257,
    "args": {
      "External id": 6813,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3464, "finished": true, "Ev Idx": 6812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 537, "tid": 537,
    "ts": 7429564860994.259, "dur": 0.088,
    "args": {
      "External id": 6814,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3465, "finished": true, "Ev Idx": 6813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 537, "tid": 537,
    "ts": 7429564860995.231, "dur": 8.589,
    "args": {
      "External id": 6815,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3465, "finished": true, "Ev Idx": 6814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564861095.875, "dur": 5.507,
    "args": {
      "External id": 6816,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3466, "finished": true, "Ev Idx": 6815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564861098.759, "dur": 0.773,
    "args": {
      "External id": 6817,"Record function id": 0, "finished": true, "Ev Idx": 6816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564861102.327, "dur": 2.593,
    "args": {
      "External id": 6818,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3467, "finished": true, "Ev Idx": 6817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564861103.849, "dur": 0.290,
    "args": {
      "External id": 6819,"Record function id": 0, "finished": true, "Ev Idx": 6818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 537, "tid": 537,
    "ts": 7429564861105.667, "dur": 1.540,
    "args": {
      "External id": 6820,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3468, "finished": true, "Ev Idx": 6819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564861106.433, "dur": 0.233,
    "args": {
      "External id": 6821,"Record function id": 0, "finished": true, "Ev Idx": 6820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 537, "tid": 537,
    "ts": 7429564861116.479, "dur": 48.279,
    "args": {
      "External id": 6822,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3469, "finished": true, "Ev Idx": 6821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 537, "tid": 537,
    "ts": 7429564861116.963, "dur": 3.014,
    "args": {
      "External id": 6823,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3469, "finished": true, "Ev Idx": 6822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 537, "tid": 537,
    "ts": 7429564861118.243, "dur": 1.154,
    "args": {
      "External id": 6824,"Record function id": 0, "finished": true, "Ev Idx": 6823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 537, "tid": 537,
    "ts": 7429564861118.846, "dur": 0.416,
    "args": {
      "External id": 6825,"Record function id": 0, "finished": true, "Ev Idx": 6824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 537, "tid": 537,
    "ts": 7429564861120.411, "dur": 44.213,
    "args": {
      "External id": 6826,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3470, "finished": true, "Ev Idx": 6825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 537, "tid": 537,
    "ts": 7429564861120.909, "dur": 1.802,
    "args": {
      "External id": 6827,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3470, "finished": true, "Ev Idx": 6826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 537, "tid": 537,
    "ts": 7429564861121.299, "dur": 1.305,
    "args": {
      "External id": 6828,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3470, "finished": true, "Ev Idx": 6827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 537, "tid": 537,
    "ts": 7429564861123.064, "dur": 39.210,
    "args": {
      "External id": 6829,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3471, "finished": true, "Ev Idx": 6828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 537, "tid": 537,
    "ts": 7429564861133.550, "dur": 1.387,
    "args": {
      "External id": 6830,"Record function id": 0, "finished": true, "Ev Idx": 6829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 537, "tid": 537,
    "ts": 7429564861162.931, "dur": 1.251,
    "args": {
      "External id": 6831,"Record function id": 0, "Fwd thread id": 0, "Sequence number": 3472, "finished": true, "Ev Idx": 6830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(1032): _bootstrap", "pid": 0, "tid": 0,
    "ts": 7429564820456.438, "dur": 41416.477,
    "args": {
      "Python parent id": null, "Python id": 1, "finished": true, "Ev Idx": 6831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(1075): _bootstrap_inner", "pid": 0, "tid": 0,
    "ts": 7429564820458.061, "dur": 41414.854,
    "args": {
      "Python parent id": 1, "Python id": 2, "finished": true, "Ev Idx": 6832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "tqdm/_monitor.py(60): run", "pid": 0, "tid": 0,
    "ts": 7429564820460.034, "dur": 41412.881,
    "args": {
      "Python parent id": 2, "Python id": 3, "finished": true, "Ev Idx": 6833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(655): wait", "pid": 0, "tid": 0,
    "ts": 7429564820460.847, "dur": 41412.068,
    "args": {
      "Python parent id": 3, "Python id": 4, "finished": true, "Ev Idx": 6834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "threading.py(359): wait", "pid": 0, "tid": 0,
    "ts": 7429564820463.107, "dur": 41409.808,
    "args": {
      "Python parent id": 4, "Python id": 5, "finished": true, "Ev Idx": 6835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "run_llama.py(20): <module>", "pid": 537, "tid": 537,
    "ts": 7429564820485.458, "dur": 41387.457,
    "args": {
      "Python parent id": null, "Python id": 6, "finished": true, "Ev Idx": 6836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(800): __enter__", "pid": 537, "tid": 537,
    "ts": 7429564820486.053, "dur": 105.830,
    "args": {
      "Python parent id": 6, "Python id": 7, "finished": true, "Ev Idx": 6837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(810): start", "pid": 537, "tid": 537,
    "ts": 7429564820486.917, "dur": 104.862,
    "args": {
      "Python parent id": 7, "Python id": 8, "finished": true, "Ev Idx": 6838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(866): _transit_action", "pid": 537, "tid": 537,
    "ts": 7429564820488.582, "dur": 102.648,
    "args": {
      "Python parent id": 8, "Python id": 9, "finished": true, "Ev Idx": 6839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(203): start_trace", "pid": 537, "tid": 537,
    "ts": 7429564820489.791, "dur": 100.958,
    "args": {
      "Python parent id": 9, "Python id": 10, "finished": true, "Ev Idx": 6840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/autograd/profiler.py(370): _start_trace", "pid": 537, "tid": 537,
    "ts": 7429564820490.332, "dur": 25.613,
    "args": {
      "Python parent id": 10, "Python id": 11, "finished": true, "Ev Idx": 6841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function perf_counter_ns>", "pid": 537, "tid": 537,
    "ts": 7429564820512.448, "dur": 1.474,
    "args": {
      "Python parent id": 11, "Python id": 12, "finished": true, "Ev Idx": 6842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(340): add_metadata_json", "pid": 537, "tid": 537,
    "ts": 7429564820520.424, "dur": 8.964,
    "args": {
      "Python parent id": 10, "Python id": 13, "finished": true, "Ev Idx": 6843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _add_metadata_json of PyCapsule object at 0x7febfd4149f0>", "pid": 537, "tid": 537,
    "ts": 7429564820526.858, "dur": 2.368,
    "args": {
      "Python parent id": 13, "Python id": 14, "finished": true, "Ev Idx": 6844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method kineto_available of PyCapsule object at 0x7febfce6f270>", "pid": 537, "tid": 537,
    "ts": 7429564820532.448, "dur": 1.074,
    "args": {
      "Python parent id": 10, "Python id": 15, "finished": true, "Ev Idx": 6845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(355): _get_distributed_info", "pid": 537, "tid": 537,
    "ts": 7429564820538.145, "dur": 14.411,
    "args": {
      "Python parent id": 10, "Python id": 16, "finished": true, "Ev Idx": 6846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/distributed/__init__.py(14): is_available", "pid": 537, "tid": 537,
    "ts": 7429564820543.680, "dur": 2.552,
    "args": {
      "Python parent id": 16, "Python id": 17, "finished": true, "Ev Idx": 6847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564820545.286, "dur": 0.783,
    "args": {
      "Python parent id": 17, "Python id": 18, "finished": true, "Ev Idx": 6848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/distributed/distributed_c10d.py(1269): is_initialized", "pid": 537, "tid": 537,
    "ts": 7429564820548.258, "dur": 4.159,
    "args": {
      "Python parent id": 16, "Python id": 19, "finished": true, "Ev Idx": 6849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/distributed/distributed_c10d.py(721): WORLD", "pid": 537, "tid": 537,
    "ts": 7429564820550.540, "dur": 1.647,
    "args": {
      "Python parent id": 19, "Python id": 20, "finished": true, "Ev Idx": 6850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/distributed/distributed_c10d.py(596): default_pg", "pid": 537, "tid": 537,
    "ts": 7429564820551.703, "dur": 0.403,
    "args": {
      "Python parent id": 20, "Python id": 21, "finished": true, "Ev Idx": 6851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564820554.492, "dur": 1.067,
    "args": {
      "Python parent id": 10, "Python id": 22, "finished": true, "Ev Idx": 6852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_config_module.py(774): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564820564.929, "dur": 21.882,
    "args": {
      "Python parent id": 10, "Python id": 23, "finished": true, "Ev Idx": 6853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_config_module.py(369): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564820569.668, "dur": 16.967,
    "args": {
      "Python parent id": 23, "Python id": 24, "finished": true, "Ev Idx": 6854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_config_module.py(425): _get_alias_val", "pid": 537, "tid": 537,
    "ts": 7429564820577.962, "dur": 4.454,
    "args": {
      "Python parent id": 24, "Python id": 25, "finished": true, "Ev Idx": 6855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_config_module.py(412): _get_alias_module_and_name", "pid": 537, "tid": 537,
    "ts": 7429564820580.933, "dur": 0.607,
    "args": {
      "Python parent id": 25, "Python id": 26, "finished": true, "Ev Idx": 6856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564820585.795, "dur": 0.615,
    "args": {
      "Python parent id": 24, "Python id": 27, "finished": true, "Ev Idx": 6857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method items of dict object at 0x7fe33db679c0>", "pid": 537, "tid": 537,
    "ts": 7429564820589.554, "dur": 0.524,
    "args": {
      "Python parent id": 10, "Python id": 28, "finished": true, "Ev Idx": 6858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/tokenization_utils_base.py(298): keys", "pid": 537, "tid": 537,
    "ts": 7429564820595.964, "dur": 1.487,
    "args": {
      "Python parent id": 6, "Python id": 29, "finished": true, "Ev Idx": 6859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method keys of dict object at 0x7fe33dc1b1c0>", "pid": 537, "tid": 537,
    "ts": 7429564820597.118, "dur": 0.270,
    "args": {
      "Python parent id": 29, "Python id": 30, "finished": true, "Ev Idx": 6860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/tokenization_utils_base.py(260): __getitem__", "pid": 537, "tid": 537,
    "ts": 7429564820601.088, "dur": 1.382,
    "args": {
      "Python parent id": 6, "Python id": 31, "finished": true, "Ev Idx": 6861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564820602.100, "dur": 0.136,
    "args": {
      "Python parent id": 31, "Python id": 32, "finished": true, "Ev Idx": 6862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/tokenization_utils_base.py(260): __getitem__", "pid": 537, "tid": 537,
    "ts": 7429564820603.374, "dur": 0.230,
    "args": {
      "Python parent id": 6, "Python id": 33, "finished": true, "Ev Idx": 6863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564820603.507, "dur": 0.055,
    "args": {
      "Python parent id": 33, "Python id": 34, "finished": true, "Ev Idx": 6864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaForCausalLM_0", "pid": 537, "tid": 537,
    "ts": 7429564820615.212, "dur": 40583.650,
    "args": {
      "Python parent id": 6, "Python module id": 0, "Python id": 35, "finished": true, "Ev Idx": 6865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564820620.544, "dur": 40577.284,
    "args": {
      "Python parent id": 35, "Python id": 36, "finished": true, "Ev Idx": 6866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564820622.386, "dur": 1.291,
    "args": {
      "Python parent id": 36, "Python id": 37, "finished": true, "Ev Idx": 6867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(949): wrapper", "pid": 537, "tid": 537,
    "ts": 7429564820628.154, "dur": 40569.261,
    "args": {
      "Python parent id": 36, "Python id": 38, "finished": true, "Ev Idx": 6868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pop of dict object at 0x7fe33db65400>", "pid": 537, "tid": 537,
    "ts": 7429564820629.778, "dur": 0.393,
    "args": {
      "Python parent id": 38, "Python id": 39, "finished": true, "Ev Idx": 6869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564820630.662, "dur": 0.299,
    "args": {
      "Python parent id": 38, "Python id": 40, "finished": true, "Ev Idx": 6870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820632.933, "dur": 6.531,
    "args": {
      "Python parent id": 38, "Python id": 41, "finished": true, "Ev Idx": 6871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(333): use_return_dict", "pid": 537, "tid": 537,
    "ts": 7429564820636.665, "dur": 2.649,
    "args": {
      "Python parent id": 41, "Python id": 42, "finished": true, "Ev Idx": 6872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820637.011, "dur": 0.917,
    "args": {
      "Python parent id": 42, "Python id": 43, "finished": true, "Ev Idx": 6873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820638.366, "dur": 0.793,
    "args": {
      "Python parent id": 42, "Python id": 44, "finished": true, "Ev Idx": 6874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564820640.248, "dur": 6.250,
    "args": {
      "Python parent id": 38, "Python id": 45, "finished": true, "Ev Idx": 6875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564820642.094, "dur": 3.690,
    "args": {
      "Python parent id": 45, "Python id": 46, "finished": true, "Ev Idx": 6876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/deprecation.py(120): wrapped_func", "pid": 537, "tid": 537,
    "ts": 7429564820651.576, "dur": 40544.993,
    "args": {
      "Python parent id": 38, "Python id": 47, "finished": true, "Ev Idx": 6877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(765): forward", "pid": 537, "tid": 537,
    "ts": 7429564820659.124, "dur": 40536.037,
    "args": {
      "Python parent id": 47, "Python id": 48, "finished": true, "Ev Idx": 6878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820659.681, "dur": 0.759,
    "args": {
      "Python parent id": 48, "Python id": 49, "finished": true, "Ev Idx": 6879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820661.035, "dur": 0.439,
    "args": {
      "Python parent id": 48, "Python id": 50, "finished": true, "Ev Idx": 6880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564820662.078, "dur": 0.678,
    "args": {
      "Python parent id": 48, "Python id": 51, "finished": true, "Ev Idx": 6881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaModel_0", "pid": 537, "tid": 537,
    "ts": 7429564820667.685, "dur": 40419.113,
    "args": {
      "Python parent id": 48, "Python module id": 0, "Python id": 52, "finished": true, "Ev Idx": 6882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564820669.827, "dur": 40415.858,
    "args": {
      "Python parent id": 52, "Python id": 53, "finished": true, "Ev Idx": 6883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564820670.076, "dur": 0.315,
    "args": {
      "Python parent id": 53, "Python id": 54, "finished": true, "Ev Idx": 6884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(949): wrapper", "pid": 537, "tid": 537,
    "ts": 7429564820673.334, "dur": 40411.728,
    "args": {
      "Python parent id": 53, "Python id": 55, "finished": true, "Ev Idx": 6885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pop of dict object at 0x7fe33db65400>", "pid": 537, "tid": 537,
    "ts": 7429564820673.719, "dur": 0.120,
    "args": {
      "Python parent id": 55, "Python id": 56, "finished": true, "Ev Idx": 6886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564820674.045, "dur": 0.190,
    "args": {
      "Python parent id": 55, "Python id": 57, "finished": true, "Ev Idx": 6887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820674.691, "dur": 2.645,
    "args": {
      "Python parent id": 55, "Python id": 58, "finished": true, "Ev Idx": 6888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(333): use_return_dict", "pid": 537, "tid": 537,
    "ts": 7429564820675.643, "dur": 1.609,
    "args": {
      "Python parent id": 58, "Python id": 59, "finished": true, "Ev Idx": 6889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820675.898, "dur": 0.579,
    "args": {
      "Python parent id": 59, "Python id": 60, "finished": true, "Ev Idx": 6890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820676.734, "dur": 0.418,
    "args": {
      "Python parent id": 59, "Python id": 61, "finished": true, "Ev Idx": 6891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564820677.618, "dur": 1.818,
    "args": {
      "Python parent id": 55, "Python id": 62, "finished": true, "Ev Idx": 6892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564820677.998, "dur": 1.314,
    "args": {
      "Python parent id": 62, "Python id": 63, "finished": true, "Ev Idx": 6893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(492): forward", "pid": 537, "tid": 537,
    "ts": 7429564820684.524, "dur": 40396.179,
    "args": {
      "Python parent id": 55, "Python id": 64, "finished": true, "Ev Idx": 6894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820690.044, "dur": 0.555,
    "args": {
      "Python parent id": 64, "Python id": 65, "finished": true, "Ev Idx": 6895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564820692.827, "dur": 0.151,
    "args": {
      "Python parent id": 64, "Python id": 66, "finished": true, "Ev Idx": 6896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564820693.981, "dur": 0.520,
    "args": {
      "Python parent id": 64, "Python id": 67, "finished": true, "Ev Idx": 6897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Embedding_0", "pid": 537, "tid": 537,
    "ts": 7429564820705.361, "dur": 176.338,
    "args": {
      "Python parent id": 64, "Python module id": 0, "Python id": 68, "finished": true, "Ev Idx": 6898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564820706.539, "dur": 174.593,
    "args": {
      "Python parent id": 68, "Python id": 69, "finished": true, "Ev Idx": 6899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564820706.791, "dur": 0.201,
    "args": {
      "Python parent id": 69, "Python id": 70, "finished": true, "Ev Idx": 6900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/sparse.py(189): forward", "pid": 537, "tid": 537,
    "ts": 7429564820710.706, "dur": 169.954,
    "args": {
      "Python parent id": 69, "Python id": 71, "finished": true, "Ev Idx": 6901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564820711.776, "dur": 0.462,
    "args": {
      "Python parent id": 71, "Python id": 72, "finished": true, "Ev Idx": 6902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2437): embedding", "pid": 537, "tid": 537,
    "ts": 7429564820715.499, "dur": 164.942,
    "args": {
      "Python parent id": 71, "Python id": 73, "finished": true, "Ev Idx": 6903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_variadic>", "pid": 537, "tid": 537,
    "ts": 7429564820716.554, "dur": 0.228,
    "args": {
      "Python parent id": 73, "Python id": 74, "finished": true, "Ev Idx": 6904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method embedding of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564820719.726, "dur": 160.524,
    "args": {
      "Python parent id": 73, "Python id": 75, "finished": true, "Ev Idx": 6905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(362): __init__", "pid": 537, "tid": 537,
    "ts": 7429564820884.986, "dur": 2.915,
    "args": {
      "Python parent id": 64, "Python id": 76, "finished": true, "Ev Idx": 6906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(30): __init__", "pid": 537, "tid": 537,
    "ts": 7429564820886.843, "dur": 0.763,
    "args": {
      "Python parent id": 76, "Python id": 77, "finished": true, "Ev Idx": 6907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(451): get_seq_length", "pid": 537, "tid": 537,
    "ts": 7429564820890.134, "dur": 1.563,
    "args": {
      "Python parent id": 64, "Python id": 78, "finished": true, "Ev Idx": 6908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564820891.163, "dur": 0.248,
    "args": {
      "Python parent id": 78, "Python id": 79, "finished": true, "Ev Idx": 6909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method arange of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564820896.017, "dur": 26.069,
    "args": {
      "Python parent id": 64, "Python id": 80, "finished": true, "Ev Idx": 6910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564820923.909, "dur": 17.563,
    "args": {
      "Python parent id": 64, "Python id": 81, "finished": true, "Ev Idx": 6911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(601): _update_causal_mask", "pid": 537, "tid": 537,
    "ts": 7429564820945.815, "dur": 318.444,
    "args": {
      "Python parent id": 64, "Python id": 82, "finished": true, "Ev Idx": 6912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820946.307, "dur": 5.877,
    "args": {
      "Python parent id": 82, "Python id": 83, "finished": true, "Ev Idx": 6913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564820948.763, "dur": 3.258,
    "args": {
      "Python parent id": 83, "Python id": 84, "finished": true, "Ev Idx": 6914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564820949.064, "dur": 1.458,
    "args": {
      "Python parent id": 84, "Python id": 85, "finished": true, "Ev Idx": 6915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820949.693, "dur": 0.736,
    "args": {
      "Python parent id": 85, "Python id": 86, "finished": true, "Ev Idx": 6916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820950.832, "dur": 0.394,
    "args": {
      "Python parent id": 84, "Python id": 87, "finished": true, "Ev Idx": 6917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820951.519, "dur": 0.424,
    "args": {
      "Python parent id": 84, "Python id": 88, "finished": true, "Ev Idx": 6918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820952.774, "dur": 2.628,
    "args": {
      "Python parent id": 82, "Python id": 89, "finished": true, "Ev Idx": 6919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564820953.384, "dur": 1.950,
    "args": {
      "Python parent id": 89, "Python id": 90, "finished": true, "Ev Idx": 6920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564820953.499, "dur": 0.612,
    "args": {
      "Python parent id": 90, "Python id": 91, "finished": true, "Ev Idx": 6921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820953.642, "dur": 0.396,
    "args": {
      "Python parent id": 91, "Python id": 92, "finished": true, "Ev Idx": 6922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820954.282, "dur": 0.398,
    "args": {
      "Python parent id": 90, "Python id": 93, "finished": true, "Ev Idx": 6923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820954.861, "dur": 0.417,
    "args": {
      "Python parent id": 90, "Python id": 94, "finished": true, "Ev Idx": 6924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(451): get_seq_length", "pid": 537, "tid": 537,
    "ts": 7429564820956.320, "dur": 0.269,
    "args": {
      "Python parent id": 82, "Python id": 95, "finished": true, "Ev Idx": 6925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564820956.482, "dur": 0.065,
    "args": {
      "Python parent id": 95, "Python id": 96, "finished": true, "Ev Idx": 6926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564820957.143, "dur": 0.305,
    "args": {
      "Python parent id": 82, "Python id": 97, "finished": true, "Ev Idx": 6927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820958.113, "dur": 2.622,
    "args": {
      "Python parent id": 82, "Python id": 98, "finished": true, "Ev Idx": 6928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564820958.727, "dur": 1.939,
    "args": {
      "Python parent id": 98, "Python id": 99, "finished": true, "Ev Idx": 6929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564820958.822, "dur": 0.685,
    "args": {
      "Python parent id": 99, "Python id": 100, "finished": true, "Ev Idx": 6930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820959.028, "dur": 0.395,
    "args": {
      "Python parent id": 100, "Python id": 101, "finished": true, "Ev Idx": 6931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820959.624, "dur": 0.376,
    "args": {
      "Python parent id": 99, "Python id": 102, "finished": true, "Ev Idx": 6932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564820960.208, "dur": 0.404,
    "args": {
      "Python parent id": 99, "Python id": 103, "finished": true, "Ev Idx": 6933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564820962.904, "dur": 0.086,
    "args": {
      "Python parent id": 82, "Python id": 104, "finished": true, "Ev Idx": 6934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(671): _prepare_4d_causal_attention_mask_with_cache_position", "pid": 537, "tid": 537,
    "ts": 7429564820967.521, "dur": 290.096,
    "args": {
      "Python parent id": 82, "Python id": 105, "finished": true, "Ev Idx": 6935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dim of Tensor object at 0x7fe33dc63ca0>", "pid": 537, "tid": 537,
    "ts": 7429564820969.079, "dur": 0.321,
    "args": {
      "Python parent id": 105, "Python id": 106, "finished": true, "Ev Idx": 6936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method full of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564820973.584, "dur": 23.808,
    "args": {
      "Python parent id": 105, "Python id": 107, "finished": true, "Ev Idx": 6937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method triu of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564820999.361, "dur": 14.982,
    "args": {
      "Python parent id": 105, "Python id": 108, "finished": true, "Ev Idx": 6938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method arange of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564821017.086, "dur": 11.659,
    "args": {
      "Python parent id": 105, "Python id": 109, "finished": true, "Ev Idx": 6939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564821030.313, "dur": 5.164,
    "args": {
      "Python parent id": 105, "Python id": 110, "finished": true, "Ev Idx": 6940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564821095.347, "dur": 9.673,
    "args": {
      "Python parent id": 105, "Python id": 111, "finished": true, "Ev Idx": 6941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method clone of Tensor object at 0x7fe33dc62ee0>", "pid": 537, "tid": 537,
    "ts": 7429564821107.050, "dur": 28.209,
    "args": {
      "Python parent id": 105, "Python id": 112, "finished": true, "Ev Idx": 6942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564821155.682, "dur": 3.654,
    "args": {
      "Python parent id": 105, "Python id": 113, "finished": true, "Ev Idx": 6943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method masked_fill of Tensor object at 0x7fe33dc631b0>", "pid": 537, "tid": 537,
    "ts": 7429564821204.160, "dur": 35.540,
    "args": {
      "Python parent id": 105, "Python id": 114, "finished": true, "Ev Idx": 6944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564821259.875, "dur": 4.253,
    "args": {
      "Python parent id": 82, "Python id": 115, "finished": true, "Ev Idx": 6945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564821261.498, "dur": 2.523,
    "args": {
      "Python parent id": 115, "Python id": 116, "finished": true, "Ev Idx": 6946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821261.653, "dur": 1.094,
    "args": {
      "Python parent id": 116, "Python id": 117, "finished": true, "Ev Idx": 6947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564821261.940, "dur": 0.684,
    "args": {
      "Python parent id": 117, "Python id": 118, "finished": true, "Ev Idx": 6948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564821262.936, "dur": 0.421,
    "args": {
      "Python parent id": 116, "Python id": 119, "finished": true, "Ev Idx": 6949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564821263.564, "dur": 0.397,
    "args": {
      "Python parent id": 116, "Python id": 120, "finished": true, "Ev Idx": 6950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821265.892, "dur": 0.725,
    "args": {
      "Python parent id": 64, "Python id": 121, "finished": true, "Ev Idx": 6951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRotaryEmbedding_0", "pid": 537, "tid": 537,
    "ts": 7429564821270.681, "dur": 320.767,
    "args": {
      "Python parent id": 64, "Python module id": 0, "Python id": 122, "finished": true, "Ev Idx": 6952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564821271.886, "dur": 319.058,
    "args": {
      "Python parent id": 122, "Python id": 123, "finished": true, "Ev Idx": 6953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564821272.457, "dur": 0.814,
    "args": {
      "Python parent id": 123, "Python id": 124, "finished": true, "Ev Idx": 6954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_contextlib.py(113): decorate_context", "pid": 537, "tid": 537,
    "ts": 7429564821275.788, "dur": 314.800,
    "args": {
      "Python parent id": 123, "Python id": 125, "finished": true, "Ev Idx": 6955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_contextlib.py(146): clone", "pid": 537, "tid": 537,
    "ts": 7429564821276.789, "dur": 8.302,
    "args": {
      "Python parent id": 125, "Python id": 126, "finished": true, "Ev Idx": 6956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/utils/_contextlib.py(154): __new__", "pid": 537, "tid": 537,
    "ts": 7429564821278.731, "dur": 1.909,
    "args": {
      "Python parent id": 126, "Python id": 127, "finished": true, "Ev Idx": 6957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method __new__ of type object at 0x556998e09300>", "pid": 537, "tid": 537,
    "ts": 7429564821280.216, "dur": 0.345,
    "args": {
      "Python parent id": 127, "Python id": 128, "finished": true, "Ev Idx": 6958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/autograd/grad_mode.py(75): __init__", "pid": 537, "tid": 537,
    "ts": 7429564821281.954, "dur": 2.970,
    "args": {
      "Python parent id": 126, "Python id": 129, "finished": true, "Ev Idx": 6959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 537, "tid": 537,
    "ts": 7429564821283.799, "dur": 0.096,
    "args": {
      "Python parent id": 129, "Python id": 130, "finished": true, "Ev Idx": 6960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/autograd/grad_mode.py(80): __enter__", "pid": 537, "tid": 537,
    "ts": 7429564821286.489, "dur": 8.867,
    "args": {
      "Python parent id": 125, "Python id": 131, "finished": true, "Ev Idx": 6961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function is_grad_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821290.310, "dur": 0.335,
    "args": {
      "Python parent id": 131, "Python id": 132, "finished": true, "Ev Idx": 6962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/autograd/grad_mode.py(184): __init__", "pid": 537, "tid": 537,
    "ts": 7429564821292.243, "dur": 2.778,
    "args": {
      "Python parent id": 131, "Python id": 133, "finished": true, "Ev Idx": 6963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function is_grad_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821292.660, "dur": 0.082,
    "args": {
      "Python parent id": 133, "Python id": 134, "finished": true, "Ev Idx": 6964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _set_grad_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821294.155, "dur": 0.745,
    "args": {
      "Python parent id": 133, "Python id": 135, "finished": true, "Ev Idx": 6965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/modeling_rope_utils.py(81): wrapper", "pid": 537, "tid": 537,
    "ts": 7429564821297.417, "dur": 288.960,
    "args": {
      "Python parent id": 125, "Python id": 136, "finished": true, "Ev Idx": 6966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(110): forward", "pid": 537, "tid": 537,
    "ts": 7429564821300.991, "dur": 282.646,
    "args": {
      "Python parent id": 136, "Python id": 137, "finished": true, "Ev Idx": 6967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821301.687, "dur": 0.714,
    "args": {
      "Python parent id": 137, "Python id": 138, "finished": true, "Ev Idx": 6968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method float of Tensor object at 0x7fe33dc62df0>", "pid": 537, "tid": 537,
    "ts": 7429564821311.296, "dur": 2.274,
    "args": {
      "Python parent id": 137, "Python id": 139, "finished": true, "Ev Idx": 6969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564821314.392, "dur": 3.784,
    "args": {
      "Python parent id": 137, "Python id": 140, "finished": true, "Ev Idx": 6970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564821318.996, "dur": 1.778,
    "args": {
      "Python parent id": 137, "Python id": 141, "finished": true, "Ev Idx": 6971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method float of Tensor object at 0x7fe33dc62df0>", "pid": 537, "tid": 537,
    "ts": 7429564821328.076, "dur": 14.974,
    "args": {
      "Python parent id": 137, "Python id": 142, "finished": true, "Ev Idx": 6972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821348.694, "dur": 0.139,
    "args": {
      "Python parent id": 137, "Python id": 143, "finished": true, "Ev Idx": 6973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/amp/autocast_mode.py(218): __init__", "pid": 537, "tid": 537,
    "ts": 7429564821368.762, "dur": 12.560,
    "args": {
      "Python parent id": 137, "Python id": 144, "finished": true, "Ev Idx": 6974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821369.144, "dur": 0.094,
    "args": {
      "Python parent id": 144, "Python id": 145, "finished": true, "Ev Idx": 6975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function get_autocast_dtype>", "pid": 537, "tid": 537,
    "ts": 7429564821370.625, "dur": 1.635,
    "args": {
      "Python parent id": 144, "Python id": 146, "finished": true, "Ev Idx": 6976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 537, "tid": 537,
    "ts": 7429564821372.986, "dur": 0.069,
    "args": {
      "Python parent id": 144, "Python id": 147, "finished": true, "Ev Idx": 6977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/amp/autocast_mode.py(28): is_autocast_available", "pid": 537, "tid": 537,
    "ts": 7429564821374.479, "dur": 1.516,
    "args": {
      "Python parent id": 144, "Python id": 148, "finished": true, "Ev Idx": 6978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _is_autocast_available>", "pid": 537, "tid": 537,
    "ts": 7429564821375.539, "dur": 0.393,
    "args": {
      "Python parent id": 148, "Python id": 149, "finished": true, "Ev Idx": 6979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _get_privateuse1_backend_name>", "pid": 537, "tid": 537,
    "ts": 7429564821377.350, "dur": 0.569,
    "args": {
      "Python parent id": 144, "Python id": 150, "finished": true, "Ev Idx": 6980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function get_autocast_dtype>", "pid": 537, "tid": 537,
    "ts": 7429564821378.384, "dur": 0.244,
    "args": {
      "Python parent id": 144, "Python id": 151, "finished": true, "Ev Idx": 6981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function is_autocast_cache_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821380.369, "dur": 0.252,
    "args": {
      "Python parent id": 144, "Python id": 152, "finished": true, "Ev Idx": 6982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/amp/autocast_mode.py(360): __enter__", "pid": 537, "tid": 537,
    "ts": 7429564821384.041, "dur": 14.259,
    "args": {
      "Python parent id": 137, "Python id": 153, "finished": true, "Ev Idx": 6983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 537, "tid": 537,
    "ts": 7429564821387.165, "dur": 0.056,
    "args": {
      "Python parent id": 153, "Python id": 154, "finished": true, "Ev Idx": 6984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function is_autocast_cache_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821387.572, "dur": 0.149,
    "args": {
      "Python parent id": 153, "Python id": 155, "finished": true, "Ev Idx": 6985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function is_autocast_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821388.782, "dur": 0.774,
    "args": {
      "Python parent id": 153, "Python id": 156, "finished": true, "Ev Idx": 6986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function get_autocast_dtype>", "pid": 537, "tid": 537,
    "ts": 7429564821389.948, "dur": 0.183,
    "args": {
      "Python parent id": 153, "Python id": 157, "finished": true, "Ev Idx": 6987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function set_autocast_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821392.177, "dur": 0.644,
    "args": {
      "Python parent id": 153, "Python id": 158, "finished": true, "Ev Idx": 6988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function set_autocast_dtype>", "pid": 537, "tid": 537,
    "ts": 7429564821393.695, "dur": 0.643,
    "args": {
      "Python parent id": 153, "Python id": 159, "finished": true, "Ev Idx": 6989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function autocast_increment_nesting>", "pid": 537, "tid": 537,
    "ts": 7429564821395.211, "dur": 0.232,
    "args": {
      "Python parent id": 153, "Python id": 160, "finished": true, "Ev Idx": 6990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function set_autocast_cache_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821396.549, "dur": 0.261,
    "args": {
      "Python parent id": 153, "Python id": 161, "finished": true, "Ev Idx": 6991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _is_torch_function_mode_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821397.988, "dur": 0.184,
    "args": {
      "Python parent id": 153, "Python id": 162, "finished": true, "Ev Idx": 6992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method float of Tensor object at 0x7fe33dc62df0>", "pid": 537, "tid": 537,
    "ts": 7429564821398.943, "dur": 1.192,
    "args": {
      "Python parent id": 137, "Python id": 163, "finished": true, "Ev Idx": 6993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method float of Tensor object at 0x7fe33dc62df0>", "pid": 537, "tid": 537,
    "ts": 7429564821400.575, "dur": 0.770,
    "args": {
      "Python parent id": 137, "Python id": 164, "finished": true, "Ev Idx": 6994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564821481.862, "dur": 5.127,
    "args": {
      "Python parent id": 137, "Python id": 165, "finished": true, "Ev Idx": 6995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564821489.203, "dur": 20.600,
    "args": {
      "Python parent id": 137, "Python id": 166, "finished": true, "Ev Idx": 6996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cos of Tensor object at 0x7fe33db6fa70>", "pid": 537, "tid": 537,
    "ts": 7429564821512.020, "dur": 14.620,
    "args": {
      "Python parent id": 137, "Python id": 167, "finished": true, "Ev Idx": 6997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method sin of Tensor object at 0x7fe33db6fa70>", "pid": 537, "tid": 537,
    "ts": 7429564821546.034, "dur": 12.196,
    "args": {
      "Python parent id": 137, "Python id": 168, "finished": true, "Ev Idx": 6998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/amp/autocast_mode.py(391): __exit__", "pid": 537, "tid": 537,
    "ts": 7429564821569.179, "dur": 9.246,
    "args": {
      "Python parent id": 137, "Python id": 169, "finished": true, "Ev Idx": 6999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 537, "tid": 537,
    "ts": 7429564821570.026, "dur": 0.095,
    "args": {
      "Python parent id": 169, "Python id": 170, "finished": true, "Ev Idx": 7000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function autocast_decrement_nesting>", "pid": 537, "tid": 537,
    "ts": 7429564821571.057, "dur": 0.275,
    "args": {
      "Python parent id": 169, "Python id": 171, "finished": true, "Ev Idx": 7001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function clear_autocast_cache>", "pid": 537, "tid": 537,
    "ts": 7429564821572.090, "dur": 0.580,
    "args": {
      "Python parent id": 169, "Python id": 172, "finished": true, "Ev Idx": 7002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function set_autocast_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821573.312, "dur": 0.720,
    "args": {
      "Python parent id": 169, "Python id": 173, "finished": true, "Ev Idx": 7003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function set_autocast_dtype>", "pid": 537, "tid": 537,
    "ts": 7429564821574.507, "dur": 0.398,
    "args": {
      "Python parent id": 169, "Python id": 174, "finished": true, "Ev Idx": 7004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function set_autocast_cache_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821575.322, "dur": 0.176,
    "args": {
      "Python parent id": 169, "Python id": 175, "finished": true, "Ev Idx": 7005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _is_torch_function_mode_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821578.174, "dur": 0.165,
    "args": {
      "Python parent id": 169, "Python id": 176, "finished": true, "Ev Idx": 7006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564821580.014, "dur": 1.800,
    "args": {
      "Python parent id": 137, "Python id": 177, "finished": true, "Ev Idx": 7007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564821582.492, "dur": 1.022,
    "args": {
      "Python parent id": 137, "Python id": 178, "finished": true, "Ev Idx": 7008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/autograd/grad_mode.py(84): __exit__", "pid": 537, "tid": 537,
    "ts": 7429564821587.968, "dur": 2.299,
    "args": {
      "Python parent id": 125, "Python id": 179, "finished": true, "Ev Idx": 7009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/autograd/grad_mode.py(184): __init__", "pid": 537, "tid": 537,
    "ts": 7429564821588.713, "dur": 1.298,
    "args": {
      "Python parent id": 179, "Python id": 180, "finished": true, "Ev Idx": 7010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function is_grad_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821588.889, "dur": 0.114,
    "args": {
      "Python parent id": 180, "Python id": 181, "finished": true, "Ev Idx": 7011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _set_grad_enabled>", "pid": 537, "tid": 537,
    "ts": 7429564821589.424, "dur": 0.469,
    "args": {
      "Python parent id": 180, "Python id": 182, "finished": true, "Ev Idx": 7012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821593.241, "dur": 0.732,
    "args": {
      "Python parent id": 64, "Python id": 183, "finished": true, "Ev Idx": 7013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564821594.926, "dur": 1.363,
    "args": {
      "Python parent id": 64, "Python id": 184, "finished": true, "Ev Idx": 7014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/container.py(319): __getitem__", "pid": 537, "tid": 537,
    "ts": 7429564821598.612, "dur": 131.299,
    "args": {
      "Python parent id": 64, "Python id": 185, "finished": true, "Ev Idx": 7015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821599.015, "dur": 0.107,
    "args": {
      "Python parent id": 185, "Python id": 186, "finished": true, "Ev Idx": 7016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of dict object at 0x7fe33db43d00>", "pid": 537, "tid": 537,
    "ts": 7429564821601.042, "dur": 0.262,
    "args": {
      "Python parent id": 185, "Python id": 187, "finished": true, "Ev Idx": 7017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/container.py(297): __init__", "pid": 537, "tid": 537,
    "ts": 7429564821605.126, "dur": 124.475,
    "args": {
      "Python parent id": 185, "Python id": 188, "finished": true, "Ev Idx": 7018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(466): __init__", "pid": 537, "tid": 537,
    "ts": 7429564821609.409, "dur": 12.240,
    "args": {
      "Python parent id": 188, "Python id": 189, "finished": true, "Ev Idx": 7019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _log_api_usage_once of PyCapsule object at 0x7fed69f96e50>", "pid": 537, "tid": 537,
    "ts": 7429564821610.968, "dur": 1.453,
    "args": {
      "Python parent id": 189, "Python id": 190, "finished": true, "Ev Idx": 7020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/container.py(348): __iadd__", "pid": 537, "tid": 537,
    "ts": 7429564821623.049, "dur": 106.315,
    "args": {
      "Python parent id": 188, "Python id": 191, "finished": true, "Ev Idx": 7021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/container.py(420): extend", "pid": 537, "tid": 537,
    "ts": 7429564821624.588, "dur": 104.627,
    "args": {
      "Python parent id": 191, "Python id": 192, "finished": true, "Ev Idx": 7022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821625.249, "dur": 3.974,
    "args": {
      "Python parent id": 192, "Python id": 193, "finished": true, "Ev Idx": 7023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen abc>(117): __instancecheck__", "pid": 537, "tid": 537,
    "ts": 7429564821626.596, "dur": 2.425,
    "args": {
      "Python parent id": 193, "Python id": 194, "finished": true, "Ev Idx": 7024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _abc_instancecheck>", "pid": 537, "tid": 537,
    "ts": 7429564821627.840, "dur": 1.080,
    "args": {
      "Python parent id": 194, "Python id": 195, "finished": true, "Ev Idx": 7025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564821629.823, "dur": 1.408,
    "args": {
      "Python parent id": 192, "Python id": 196, "finished": true, "Ev Idx": 7026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/container.py(340): __len__", "pid": 537, "tid": 537,
    "ts": 7429564821630.794, "dur": 0.343,
    "args": {
      "Python parent id": 196, "Python id": 197, "finished": true, "Ev Idx": 7027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564821631.000, "dur": 0.077,
    "args": {
      "Python parent id": 197, "Python id": 198, "finished": true, "Ev Idx": 7028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821636.538, "dur": 8.919,
    "args": {
      "Python parent id": 192, "Python id": 199, "finished": true, "Ev Idx": 7029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821636.820, "dur": 0.276,
    "args": {
      "Python parent id": 199, "Python id": 200, "finished": true, "Ev Idx": 7030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821637.520, "dur": 0.048,
    "args": {
      "Python parent id": 199, "Python id": 201, "finished": true, "Ev Idx": 7031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821637.953, "dur": 4.457,
    "args": {
      "Python parent id": 199, "Python id": 202, "finished": true, "Ev Idx": 7032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821638.808, "dur": 3.061,
    "args": {
      "Python parent id": 202, "Python id": 203, "finished": true, "Ev Idx": 7033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821644.365, "dur": 0.265,
    "args": {
      "Python parent id": 199, "Python id": 204, "finished": true, "Ev Idx": 7034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821646.661, "dur": 3.535,
    "args": {
      "Python parent id": 192, "Python id": 205, "finished": true, "Ev Idx": 7035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821646.820, "dur": 0.134,
    "args": {
      "Python parent id": 205, "Python id": 206, "finished": true, "Ev Idx": 7036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821647.111, "dur": 0.063,
    "args": {
      "Python parent id": 205, "Python id": 207, "finished": true, "Ev Idx": 7037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821647.428, "dur": 1.951,
    "args": {
      "Python parent id": 205, "Python id": 208, "finished": true, "Ev Idx": 7038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821648.140, "dur": 1.059,
    "args": {
      "Python parent id": 208, "Python id": 209, "finished": true, "Ev Idx": 7039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821649.788, "dur": 0.103,
    "args": {
      "Python parent id": 205, "Python id": 210, "finished": true, "Ev Idx": 7040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821650.743, "dur": 2.227,
    "args": {
      "Python parent id": 192, "Python id": 211, "finished": true, "Ev Idx": 7041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821650.847, "dur": 0.046,
    "args": {
      "Python parent id": 211, "Python id": 212, "finished": true, "Ev Idx": 7042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821651.001, "dur": 0.041,
    "args": {
      "Python parent id": 211, "Python id": 213, "finished": true, "Ev Idx": 7043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821651.192, "dur": 1.286,
    "args": {
      "Python parent id": 211, "Python id": 214, "finished": true, "Ev Idx": 7044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821651.725, "dur": 0.636,
    "args": {
      "Python parent id": 214, "Python id": 215, "finished": true, "Ev Idx": 7045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821652.735, "dur": 0.072,
    "args": {
      "Python parent id": 211, "Python id": 216, "finished": true, "Ev Idx": 7046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821653.380, "dur": 2.629,
    "args": {
      "Python parent id": 192, "Python id": 217, "finished": true, "Ev Idx": 7047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821653.483, "dur": 0.038,
    "args": {
      "Python parent id": 217, "Python id": 218, "finished": true, "Ev Idx": 7048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821653.630, "dur": 0.030,
    "args": {
      "Python parent id": 217, "Python id": 219, "finished": true, "Ev Idx": 7049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821653.815, "dur": 1.637,
    "args": {
      "Python parent id": 217, "Python id": 220, "finished": true, "Ev Idx": 7050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821654.566, "dur": 0.566,
    "args": {
      "Python parent id": 220, "Python id": 221, "finished": true, "Ev Idx": 7051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821655.780, "dur": 0.083,
    "args": {
      "Python parent id": 217, "Python id": 222, "finished": true, "Ev Idx": 7052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821656.560, "dur": 1.920,
    "args": {
      "Python parent id": 192, "Python id": 223, "finished": true, "Ev Idx": 7053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821656.631, "dur": 0.050,
    "args": {
      "Python parent id": 223, "Python id": 224, "finished": true, "Ev Idx": 7054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821656.778, "dur": 0.035,
    "args": {
      "Python parent id": 223, "Python id": 225, "finished": true, "Ev Idx": 7055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821656.946, "dur": 1.023,
    "args": {
      "Python parent id": 223, "Python id": 226, "finished": true, "Ev Idx": 7056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821657.405, "dur": 0.458,
    "args": {
      "Python parent id": 226, "Python id": 227, "finished": true, "Ev Idx": 7057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821658.235, "dur": 0.093,
    "args": {
      "Python parent id": 223, "Python id": 228, "finished": true, "Ev Idx": 7058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821658.974, "dur": 2.520,
    "args": {
      "Python parent id": 192, "Python id": 229, "finished": true, "Ev Idx": 7059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821659.071, "dur": 0.048,
    "args": {
      "Python parent id": 229, "Python id": 230, "finished": true, "Ev Idx": 7060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821659.332, "dur": 0.035,
    "args": {
      "Python parent id": 229, "Python id": 231, "finished": true, "Ev Idx": 7061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821659.502, "dur": 1.166,
    "args": {
      "Python parent id": 229, "Python id": 232, "finished": true, "Ev Idx": 7062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821659.993, "dur": 0.572,
    "args": {
      "Python parent id": 232, "Python id": 233, "finished": true, "Ev Idx": 7063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821660.938, "dur": 0.196,
    "args": {
      "Python parent id": 229, "Python id": 234, "finished": true, "Ev Idx": 7064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821661.994, "dur": 1.908,
    "args": {
      "Python parent id": 192, "Python id": 235, "finished": true, "Ev Idx": 7065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821662.099, "dur": 0.044,
    "args": {
      "Python parent id": 235, "Python id": 236, "finished": true, "Ev Idx": 7066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821662.271, "dur": 0.040,
    "args": {
      "Python parent id": 235, "Python id": 237, "finished": true, "Ev Idx": 7067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821662.452, "dur": 0.949,
    "args": {
      "Python parent id": 235, "Python id": 238, "finished": true, "Ev Idx": 7068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821662.763, "dur": 0.536,
    "args": {
      "Python parent id": 238, "Python id": 239, "finished": true, "Ev Idx": 7069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821663.671, "dur": 0.074,
    "args": {
      "Python parent id": 235, "Python id": 240, "finished": true, "Ev Idx": 7070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821664.325, "dur": 2.283,
    "args": {
      "Python parent id": 192, "Python id": 241, "finished": true, "Ev Idx": 7071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821664.417, "dur": 0.045,
    "args": {
      "Python parent id": 241, "Python id": 242, "finished": true, "Ev Idx": 7072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821664.691, "dur": 0.038,
    "args": {
      "Python parent id": 241, "Python id": 243, "finished": true, "Ev Idx": 7073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821664.882, "dur": 1.112,
    "args": {
      "Python parent id": 241, "Python id": 244, "finished": true, "Ev Idx": 7074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821665.253, "dur": 0.504,
    "args": {
      "Python parent id": 244, "Python id": 245, "finished": true, "Ev Idx": 7075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821666.249, "dur": 0.140,
    "args": {
      "Python parent id": 241, "Python id": 246, "finished": true, "Ev Idx": 7076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821667.192, "dur": 2.046,
    "args": {
      "Python parent id": 192, "Python id": 247, "finished": true, "Ev Idx": 7077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821667.267, "dur": 0.042,
    "args": {
      "Python parent id": 247, "Python id": 248, "finished": true, "Ev Idx": 7078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821667.412, "dur": 0.036,
    "args": {
      "Python parent id": 247, "Python id": 249, "finished": true, "Ev Idx": 7079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821667.579, "dur": 1.068,
    "args": {
      "Python parent id": 247, "Python id": 250, "finished": true, "Ev Idx": 7080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821667.969, "dur": 0.572,
    "args": {
      "Python parent id": 250, "Python id": 251, "finished": true, "Ev Idx": 7081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821668.914, "dur": 0.094,
    "args": {
      "Python parent id": 247, "Python id": 252, "finished": true, "Ev Idx": 7082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821669.720, "dur": 2.003,
    "args": {
      "Python parent id": 192, "Python id": 253, "finished": true, "Ev Idx": 7083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821669.789, "dur": 0.054,
    "args": {
      "Python parent id": 253, "Python id": 254, "finished": true, "Ev Idx": 7084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821669.982, "dur": 0.031,
    "args": {
      "Python parent id": 253, "Python id": 255, "finished": true, "Ev Idx": 7085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821670.151, "dur": 1.018,
    "args": {
      "Python parent id": 253, "Python id": 256, "finished": true, "Ev Idx": 7086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821670.504, "dur": 0.538,
    "args": {
      "Python parent id": 256, "Python id": 257, "finished": true, "Ev Idx": 7087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821671.441, "dur": 0.076,
    "args": {
      "Python parent id": 253, "Python id": 258, "finished": true, "Ev Idx": 7088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821672.301, "dur": 2.634,
    "args": {
      "Python parent id": 192, "Python id": 259, "finished": true, "Ev Idx": 7089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821672.373, "dur": 0.047,
    "args": {
      "Python parent id": 259, "Python id": 260, "finished": true, "Ev Idx": 7090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821672.531, "dur": 0.038,
    "args": {
      "Python parent id": 259, "Python id": 261, "finished": true, "Ev Idx": 7091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821672.712, "dur": 1.079,
    "args": {
      "Python parent id": 259, "Python id": 262, "finished": true, "Ev Idx": 7092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821673.079, "dur": 0.610,
    "args": {
      "Python parent id": 262, "Python id": 263, "finished": true, "Ev Idx": 7093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821674.122, "dur": 0.081,
    "args": {
      "Python parent id": 259, "Python id": 264, "finished": true, "Ev Idx": 7094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821675.412, "dur": 1.970,
    "args": {
      "Python parent id": 192, "Python id": 265, "finished": true, "Ev Idx": 7095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821675.482, "dur": 0.045,
    "args": {
      "Python parent id": 265, "Python id": 266, "finished": true, "Ev Idx": 7096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821675.631, "dur": 0.031,
    "args": {
      "Python parent id": 265, "Python id": 267, "finished": true, "Ev Idx": 7097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821675.795, "dur": 1.095,
    "args": {
      "Python parent id": 265, "Python id": 268, "finished": true, "Ev Idx": 7098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821676.211, "dur": 0.580,
    "args": {
      "Python parent id": 268, "Python id": 269, "finished": true, "Ev Idx": 7099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821677.167, "dur": 0.071,
    "args": {
      "Python parent id": 265, "Python id": 270, "finished": true, "Ev Idx": 7100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821677.792, "dur": 2.078,
    "args": {
      "Python parent id": 192, "Python id": 271, "finished": true, "Ev Idx": 7101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821677.864, "dur": 0.043,
    "args": {
      "Python parent id": 271, "Python id": 272, "finished": true, "Ev Idx": 7102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821678.054, "dur": 0.039,
    "args": {
      "Python parent id": 271, "Python id": 273, "finished": true, "Ev Idx": 7103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821678.240, "dur": 0.979,
    "args": {
      "Python parent id": 271, "Python id": 274, "finished": true, "Ev Idx": 7104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821678.600, "dur": 0.514,
    "args": {
      "Python parent id": 274, "Python id": 275, "finished": true, "Ev Idx": 7105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821679.518, "dur": 0.101,
    "args": {
      "Python parent id": 271, "Python id": 276, "finished": true, "Ev Idx": 7106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821680.292, "dur": 2.080,
    "args": {
      "Python parent id": 192, "Python id": 277, "finished": true, "Ev Idx": 7107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821680.363, "dur": 0.050,
    "args": {
      "Python parent id": 277, "Python id": 278, "finished": true, "Ev Idx": 7108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821680.518, "dur": 0.037,
    "args": {
      "Python parent id": 277, "Python id": 279, "finished": true, "Ev Idx": 7109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821680.694, "dur": 1.158,
    "args": {
      "Python parent id": 277, "Python id": 280, "finished": true, "Ev Idx": 7110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821681.223, "dur": 0.523,
    "args": {
      "Python parent id": 280, "Python id": 281, "finished": true, "Ev Idx": 7111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821682.097, "dur": 0.090,
    "args": {
      "Python parent id": 277, "Python id": 282, "finished": true, "Ev Idx": 7112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821682.798, "dur": 1.815,
    "args": {
      "Python parent id": 192, "Python id": 283, "finished": true, "Ev Idx": 7113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821682.901, "dur": 0.049,
    "args": {
      "Python parent id": 283, "Python id": 284, "finished": true, "Ev Idx": 7114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821683.057, "dur": 0.034,
    "args": {
      "Python parent id": 283, "Python id": 285, "finished": true, "Ev Idx": 7115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821683.223, "dur": 0.887,
    "args": {
      "Python parent id": 283, "Python id": 286, "finished": true, "Ev Idx": 7116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821683.515, "dur": 0.492,
    "args": {
      "Python parent id": 286, "Python id": 287, "finished": true, "Ev Idx": 7117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821684.383, "dur": 0.092,
    "args": {
      "Python parent id": 283, "Python id": 288, "finished": true, "Ev Idx": 7118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821685.015, "dur": 2.239,
    "args": {
      "Python parent id": 192, "Python id": 289, "finished": true, "Ev Idx": 7119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821685.086, "dur": 0.045,
    "args": {
      "Python parent id": 289, "Python id": 290, "finished": true, "Ev Idx": 7120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821685.228, "dur": 0.033,
    "args": {
      "Python parent id": 289, "Python id": 291, "finished": true, "Ev Idx": 7121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821685.392, "dur": 1.353,
    "args": {
      "Python parent id": 289, "Python id": 292, "finished": true, "Ev Idx": 7122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821686.089, "dur": 0.549,
    "args": {
      "Python parent id": 292, "Python id": 293, "finished": true, "Ev Idx": 7123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821686.996, "dur": 0.102,
    "args": {
      "Python parent id": 289, "Python id": 294, "finished": true, "Ev Idx": 7124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821687.699, "dur": 1.890,
    "args": {
      "Python parent id": 192, "Python id": 295, "finished": true, "Ev Idx": 7125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821687.772, "dur": 0.043,
    "args": {
      "Python parent id": 295, "Python id": 296, "finished": true, "Ev Idx": 7126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821687.936, "dur": 0.032,
    "args": {
      "Python parent id": 295, "Python id": 297, "finished": true, "Ev Idx": 7127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821688.132, "dur": 1.018,
    "args": {
      "Python parent id": 295, "Python id": 298, "finished": true, "Ev Idx": 7128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821688.540, "dur": 0.486,
    "args": {
      "Python parent id": 298, "Python id": 299, "finished": true, "Ev Idx": 7129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821689.395, "dur": 0.077,
    "args": {
      "Python parent id": 295, "Python id": 300, "finished": true, "Ev Idx": 7130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821690.014, "dur": 2.646,
    "args": {
      "Python parent id": 192, "Python id": 301, "finished": true, "Ev Idx": 7131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821690.107, "dur": 0.048,
    "args": {
      "Python parent id": 301, "Python id": 302, "finished": true, "Ev Idx": 7132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821690.260, "dur": 0.037,
    "args": {
      "Python parent id": 301, "Python id": 303, "finished": true, "Ev Idx": 7133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821690.450, "dur": 1.105,
    "args": {
      "Python parent id": 301, "Python id": 304, "finished": true, "Ev Idx": 7134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821690.968, "dur": 0.454,
    "args": {
      "Python parent id": 304, "Python id": 305, "finished": true, "Ev Idx": 7135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821691.800, "dur": 0.198,
    "args": {
      "Python parent id": 301, "Python id": 306, "finished": true, "Ev Idx": 7136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821693.138, "dur": 2.278,
    "args": {
      "Python parent id": 192, "Python id": 307, "finished": true, "Ev Idx": 7137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821693.214, "dur": 0.050,
    "args": {
      "Python parent id": 307, "Python id": 308, "finished": true, "Ev Idx": 7138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821693.375, "dur": 0.039,
    "args": {
      "Python parent id": 307, "Python id": 309, "finished": true, "Ev Idx": 7139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821693.550, "dur": 1.342,
    "args": {
      "Python parent id": 307, "Python id": 310, "finished": true, "Ev Idx": 7140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821694.193, "dur": 0.591,
    "args": {
      "Python parent id": 310, "Python id": 311, "finished": true, "Ev Idx": 7141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821695.142, "dur": 0.097,
    "args": {
      "Python parent id": 307, "Python id": 312, "finished": true, "Ev Idx": 7142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821695.857, "dur": 2.168,
    "args": {
      "Python parent id": 192, "Python id": 313, "finished": true, "Ev Idx": 7143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821695.930, "dur": 0.049,
    "args": {
      "Python parent id": 313, "Python id": 314, "finished": true, "Ev Idx": 7144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821696.080, "dur": 0.034,
    "args": {
      "Python parent id": 313, "Python id": 315, "finished": true, "Ev Idx": 7145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821696.248, "dur": 1.142,
    "args": {
      "Python parent id": 313, "Python id": 316, "finished": true, "Ev Idx": 7146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821696.717, "dur": 0.564,
    "args": {
      "Python parent id": 316, "Python id": 317, "finished": true, "Ev Idx": 7147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821697.657, "dur": 0.101,
    "args": {
      "Python parent id": 313, "Python id": 318, "finished": true, "Ev Idx": 7148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821698.477, "dur": 1.941,
    "args": {
      "Python parent id": 192, "Python id": 319, "finished": true, "Ev Idx": 7149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821698.575, "dur": 0.042,
    "args": {
      "Python parent id": 319, "Python id": 320, "finished": true, "Ev Idx": 7150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821698.715, "dur": 0.034,
    "args": {
      "Python parent id": 319, "Python id": 321, "finished": true, "Ev Idx": 7151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821698.883, "dur": 0.975,
    "args": {
      "Python parent id": 319, "Python id": 322, "finished": true, "Ev Idx": 7152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821699.215, "dur": 0.540,
    "args": {
      "Python parent id": 322, "Python id": 323, "finished": true, "Ev Idx": 7153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821700.130, "dur": 0.089,
    "args": {
      "Python parent id": 319, "Python id": 324, "finished": true, "Ev Idx": 7154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821700.829, "dur": 2.565,
    "args": {
      "Python parent id": 192, "Python id": 325, "finished": true, "Ev Idx": 7155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821700.900, "dur": 0.045,
    "args": {
      "Python parent id": 325, "Python id": 326, "finished": true, "Ev Idx": 7156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821701.043, "dur": 0.034,
    "args": {
      "Python parent id": 325, "Python id": 327, "finished": true, "Ev Idx": 7157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821701.206, "dur": 1.344,
    "args": {
      "Python parent id": 325, "Python id": 328, "finished": true, "Ev Idx": 7158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821701.930, "dur": 0.511,
    "args": {
      "Python parent id": 328, "Python id": 329, "finished": true, "Ev Idx": 7159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821702.804, "dur": 0.085,
    "args": {
      "Python parent id": 325, "Python id": 330, "finished": true, "Ev Idx": 7160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821703.785, "dur": 1.992,
    "args": {
      "Python parent id": 192, "Python id": 331, "finished": true, "Ev Idx": 7161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821703.859, "dur": 0.048,
    "args": {
      "Python parent id": 331, "Python id": 332, "finished": true, "Ev Idx": 7162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821704.028, "dur": 0.032,
    "args": {
      "Python parent id": 331, "Python id": 333, "finished": true, "Ev Idx": 7163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821704.193, "dur": 1.061,
    "args": {
      "Python parent id": 331, "Python id": 334, "finished": true, "Ev Idx": 7164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821704.675, "dur": 0.469,
    "args": {
      "Python parent id": 334, "Python id": 335, "finished": true, "Ev Idx": 7165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821705.526, "dur": 0.098,
    "args": {
      "Python parent id": 331, "Python id": 336, "finished": true, "Ev Idx": 7166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821706.151, "dur": 2.357,
    "args": {
      "Python parent id": 192, "Python id": 337, "finished": true, "Ev Idx": 7167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821706.220, "dur": 0.045,
    "args": {
      "Python parent id": 337, "Python id": 338, "finished": true, "Ev Idx": 7168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821706.365, "dur": 0.034,
    "args": {
      "Python parent id": 337, "Python id": 339, "finished": true, "Ev Idx": 7169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821706.557, "dur": 1.368,
    "args": {
      "Python parent id": 337, "Python id": 340, "finished": true, "Ev Idx": 7170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821707.191, "dur": 0.496,
    "args": {
      "Python parent id": 340, "Python id": 341, "finished": true, "Ev Idx": 7171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821708.196, "dur": 0.182,
    "args": {
      "Python parent id": 337, "Python id": 342, "finished": true, "Ev Idx": 7172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821708.922, "dur": 2.113,
    "args": {
      "Python parent id": 192, "Python id": 343, "finished": true, "Ev Idx": 7173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821708.992, "dur": 0.041,
    "args": {
      "Python parent id": 343, "Python id": 344, "finished": true, "Ev Idx": 7174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821709.131, "dur": 0.039,
    "args": {
      "Python parent id": 343, "Python id": 345, "finished": true, "Ev Idx": 7175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821709.300, "dur": 1.192,
    "args": {
      "Python parent id": 343, "Python id": 346, "finished": true, "Ev Idx": 7176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821709.823, "dur": 0.567,
    "args": {
      "Python parent id": 346, "Python id": 347, "finished": true, "Ev Idx": 7177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821710.773, "dur": 0.089,
    "args": {
      "Python parent id": 343, "Python id": 348, "finished": true, "Ev Idx": 7178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821711.478, "dur": 2.465,
    "args": {
      "Python parent id": 192, "Python id": 349, "finished": true, "Ev Idx": 7179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821711.552, "dur": 0.045,
    "args": {
      "Python parent id": 349, "Python id": 350, "finished": true, "Ev Idx": 7180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821711.721, "dur": 0.061,
    "args": {
      "Python parent id": 349, "Python id": 351, "finished": true, "Ev Idx": 7181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821711.918, "dur": 1.102,
    "args": {
      "Python parent id": 349, "Python id": 352, "finished": true, "Ev Idx": 7182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821712.405, "dur": 0.498,
    "args": {
      "Python parent id": 352, "Python id": 353, "finished": true, "Ev Idx": 7183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821713.465, "dur": 0.175,
    "args": {
      "Python parent id": 349, "Python id": 354, "finished": true, "Ev Idx": 7184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821714.400, "dur": 1.992,
    "args": {
      "Python parent id": 192, "Python id": 355, "finished": true, "Ev Idx": 7185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821714.470, "dur": 0.046,
    "args": {
      "Python parent id": 355, "Python id": 356, "finished": true, "Ev Idx": 7186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821714.615, "dur": 0.033,
    "args": {
      "Python parent id": 355, "Python id": 357, "finished": true, "Ev Idx": 7187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821714.779, "dur": 1.093,
    "args": {
      "Python parent id": 355, "Python id": 358, "finished": true, "Ev Idx": 7188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821715.285, "dur": 0.482,
    "args": {
      "Python parent id": 358, "Python id": 359, "finished": true, "Ev Idx": 7189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821716.140, "dur": 0.102,
    "args": {
      "Python parent id": 355, "Python id": 360, "finished": true, "Ev Idx": 7190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821716.829, "dur": 2.349,
    "args": {
      "Python parent id": 192, "Python id": 361, "finished": true, "Ev Idx": 7191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821716.899, "dur": 0.069,
    "args": {
      "Python parent id": 361, "Python id": 362, "finished": true, "Ev Idx": 7192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821717.140, "dur": 0.032,
    "args": {
      "Python parent id": 361, "Python id": 363, "finished": true, "Ev Idx": 7193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821717.328, "dur": 1.351,
    "args": {
      "Python parent id": 361, "Python id": 364, "finished": true, "Ev Idx": 7194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821717.882, "dur": 0.566,
    "args": {
      "Python parent id": 364, "Python id": 365, "finished": true, "Ev Idx": 7195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821718.945, "dur": 0.092,
    "args": {
      "Python parent id": 361, "Python id": 366, "finished": true, "Ev Idx": 7196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821719.550, "dur": 1.834,
    "args": {
      "Python parent id": 192, "Python id": 367, "finished": true, "Ev Idx": 7197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821719.622, "dur": 0.043,
    "args": {
      "Python parent id": 367, "Python id": 368, "finished": true, "Ev Idx": 7198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821719.766, "dur": 0.036,
    "args": {
      "Python parent id": 367, "Python id": 369, "finished": true, "Ev Idx": 7199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821719.937, "dur": 0.934,
    "args": {
      "Python parent id": 367, "Python id": 370, "finished": true, "Ev Idx": 7200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821720.284, "dur": 0.481,
    "args": {
      "Python parent id": 370, "Python id": 371, "finished": true, "Ev Idx": 7201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821721.123, "dur": 0.102,
    "args": {
      "Python parent id": 367, "Python id": 372, "finished": true, "Ev Idx": 7202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821721.845, "dur": 2.015,
    "args": {
      "Python parent id": 192, "Python id": 373, "finished": true, "Ev Idx": 7203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821721.914, "dur": 0.047,
    "args": {
      "Python parent id": 373, "Python id": 374, "finished": true, "Ev Idx": 7204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821722.064, "dur": 0.033,
    "args": {
      "Python parent id": 373, "Python id": 375, "finished": true, "Ev Idx": 7205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821722.225, "dur": 1.125,
    "args": {
      "Python parent id": 373, "Python id": 376, "finished": true, "Ev Idx": 7206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821722.768, "dur": 0.481,
    "args": {
      "Python parent id": 376, "Python id": 377, "finished": true, "Ev Idx": 7207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821723.606, "dur": 0.094,
    "args": {
      "Python parent id": 373, "Python id": 378, "finished": true, "Ev Idx": 7208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821724.267, "dur": 1.944,
    "args": {
      "Python parent id": 192, "Python id": 379, "finished": true, "Ev Idx": 7209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821724.340, "dur": 0.044,
    "args": {
      "Python parent id": 379, "Python id": 380, "finished": true, "Ev Idx": 7210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821724.488, "dur": 0.032,
    "args": {
      "Python parent id": 379, "Python id": 381, "finished": true, "Ev Idx": 7211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821724.653, "dur": 0.999,
    "args": {
      "Python parent id": 379, "Python id": 382, "finished": true, "Ev Idx": 7212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821725.055, "dur": 0.498,
    "args": {
      "Python parent id": 382, "Python id": 383, "finished": true, "Ev Idx": 7213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821725.915, "dur": 0.117,
    "args": {
      "Python parent id": 379, "Python id": 384, "finished": true, "Ev Idx": 7214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(624): add_module", "pid": 537, "tid": 537,
    "ts": 7429564821726.752, "dur": 2.066,
    "args": {
      "Python parent id": 192, "Python id": 385, "finished": true, "Ev Idx": 7215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821726.827, "dur": 0.042,
    "args": {
      "Python parent id": 385, "Python id": 386, "finished": true, "Ev Idx": 7216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564821726.970, "dur": 0.035,
    "args": {
      "Python parent id": 385, "Python id": 387, "finished": true, "Ev Idx": 7217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564821727.134, "dur": 1.036,
    "args": {
      "Python parent id": 385, "Python id": 388, "finished": true, "Ev Idx": 7218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821727.531, "dur": 0.523,
    "args": {
      "Python parent id": 388, "Python id": 389, "finished": true, "Ev Idx": 7219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of collections.OrderedDict object at 0x7febfd1d4d40>", "pid": 537, "tid": 537,
    "ts": 7429564821728.423, "dur": 0.230,
    "args": {
      "Python parent id": 385, "Python id": 390, "finished": true, "Ev Idx": 7220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/container.py(344): __iter__", "pid": 537, "tid": 537,
    "ts": 7429564821731.257, "dur": 1.882,
    "args": {
      "Python parent id": 64, "Python id": 391, "finished": true, "Ev Idx": 7221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of dict object at 0x7fe33db43d00>", "pid": 537, "tid": 537,
    "ts": 7429564821731.960, "dur": 0.175,
    "args": {
      "Python parent id": 391, "Python id": 392, "finished": true, "Ev Idx": 7222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function iter>", "pid": 537, "tid": 537,
    "ts": 7429564821732.902, "dur": 0.124,
    "args": {
      "Python parent id": 391, "Python id": 393, "finished": true, "Ev Idx": 7223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_0", "pid": 537, "tid": 537,
    "ts": 7429564821741.845, "dur": 1324.259,
    "args": {
      "Python parent id": 64, "Python module id": 0, "Python id": 394, "finished": true, "Ev Idx": 7224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564821743.815, "dur": 1321.247,
    "args": {
      "Python parent id": 394, "Python id": 395, "finished": true, "Ev Idx": 7225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564821744.163, "dur": 0.820,
    "args": {
      "Python parent id": 395, "Python id": 396, "finished": true, "Ev Idx": 7226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564821748.495, "dur": 1315.338,
    "args": {
      "Python parent id": 395, "Python id": 397, "finished": true, "Ev Idx": 7227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821749.047, "dur": 0.846,
    "args": {
      "Python parent id": 397, "Python id": 398, "finished": true, "Ev Idx": 7228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_0", "pid": 537, "tid": 537,
    "ts": 7429564821755.941, "dur": 125.310,
    "args": {
      "Python parent id": 397, "Python module id": 0, "Python id": 399, "finished": true, "Ev Idx": 7229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564821757.402, "dur": 123.334,
    "args": {
      "Python parent id": 399, "Python id": 400, "finished": true, "Ev Idx": 7230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564821757.567, "dur": 0.253,
    "args": {
      "Python parent id": 400, "Python id": 401, "finished": true, "Ev Idx": 7231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564821760.370, "dur": 119.129,
    "args": {
      "Python parent id": 400, "Python id": 402, "finished": true, "Ev Idx": 7232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564821761.168, "dur": 2.714,
    "args": {
      "Python parent id": 402, "Python id": 403, "finished": true, "Ev Idx": 7233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564821765.306, "dur": 25.635,
    "args": {
      "Python parent id": 402, "Python id": 404, "finished": true, "Ev Idx": 7234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564821792.229, "dur": 26.373,
    "args": {
      "Python parent id": 402, "Python id": 405, "finished": true, "Ev Idx": 7235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564821838.789, "dur": 14.916,
    "args": {
      "Python parent id": 402, "Python id": 406, "finished": true, "Ev Idx": 7236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821867.097, "dur": 0.367,
    "args": {
      "Python parent id": 402, "Python id": 407, "finished": true, "Ev Idx": 7237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564821868.046, "dur": 1.692,
    "args": {
      "Python parent id": 402, "Python id": 408, "finished": true, "Ev Idx": 7238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821882.420, "dur": 0.561,
    "args": {
      "Python parent id": 397, "Python id": 409, "finished": true, "Ev Idx": 7239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_0", "pid": 537, "tid": 537,
    "ts": 7429564821887.078, "dur": 849.070,
    "args": {
      "Python parent id": 397, "Python module id": 0, "Python id": 410, "finished": true, "Ev Idx": 7240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564821889.022, "dur": 846.512,
    "args": {
      "Python parent id": 410, "Python id": 411, "finished": true, "Ev Idx": 7241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564821889.301, "dur": 0.348,
    "args": {
      "Python parent id": 411, "Python id": 412, "finished": true, "Ev Idx": 7242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564821894.694, "dur": 839.846,
    "args": {
      "Python parent id": 411, "Python id": 413, "finished": true, "Ev Idx": 7243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821898.122, "dur": 0.370,
    "args": {
      "Python parent id": 413, "Python id": 414, "finished": true, "Ev Idx": 7244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_0", "pid": 537, "tid": 537,
    "ts": 7429564821902.200, "dur": 91.812,
    "args": {
      "Python parent id": 413, "Python module id": 0, "Python id": 415, "finished": true, "Ev Idx": 7245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564821903.313, "dur": 90.348,
    "args": {
      "Python parent id": 415, "Python id": 416, "finished": true, "Ev Idx": 7246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564821903.487, "dur": 0.174,
    "args": {
      "Python parent id": 416, "Python id": 417, "finished": true, "Ev Idx": 7247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564821905.271, "dur": 88.076,
    "args": {
      "Python parent id": 416, "Python id": 418, "finished": true, "Ev Idx": 7248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821906.457, "dur": 0.157,
    "args": {
      "Python parent id": 418, "Python id": 419, "finished": true, "Ev Idx": 7249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564821907.091, "dur": 0.183,
    "args": {
      "Python parent id": 418, "Python id": 420, "finished": true, "Ev Idx": 7250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564821908.275, "dur": 84.924,
    "args": {
      "Python parent id": 418, "Python id": 421, "finished": true, "Ev Idx": 7251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564821995.779, "dur": 4.244,
    "args": {
      "Python parent id": 413, "Python id": 422, "finished": true, "Ev Idx": 7252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564822000.610, "dur": 4.658,
    "args": {
      "Python parent id": 413, "Python id": 423, "finished": true, "Ev Idx": 7253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822006.552, "dur": 0.593,
    "args": {
      "Python parent id": 413, "Python id": 424, "finished": true, "Ev Idx": 7254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_1", "pid": 537, "tid": 537,
    "ts": 7429564822010.843, "dur": 91.907,
    "args": {
      "Python parent id": 413, "Python module id": 1, "Python id": 425, "finished": true, "Ev Idx": 7255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564822011.430, "dur": 91.015,
    "args": {
      "Python parent id": 425, "Python id": 426, "finished": true, "Ev Idx": 7256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564822011.681, "dur": 0.362,
    "args": {
      "Python parent id": 426, "Python id": 427, "finished": true, "Ev Idx": 7257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564822013.075, "dur": 89.001,
    "args": {
      "Python parent id": 426, "Python id": 428, "finished": true, "Ev Idx": 7258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822013.358, "dur": 0.171,
    "args": {
      "Python parent id": 428, "Python id": 429, "finished": true, "Ev Idx": 7259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822013.787, "dur": 0.154,
    "args": {
      "Python parent id": 428, "Python id": 430, "finished": true, "Ev Idx": 7260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564822014.133, "dur": 87.808,
    "args": {
      "Python parent id": 428, "Python id": 431, "finished": true, "Ev Idx": 7261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564822103.526, "dur": 4.420,
    "args": {
      "Python parent id": 413, "Python id": 432, "finished": true, "Ev Idx": 7262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564822108.668, "dur": 4.175,
    "args": {
      "Python parent id": 413, "Python id": 433, "finished": true, "Ev Idx": 7263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822114.237, "dur": 0.644,
    "args": {
      "Python parent id": 413, "Python id": 434, "finished": true, "Ev Idx": 7264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_2", "pid": 537, "tid": 537,
    "ts": 7429564822117.771, "dur": 53.729,
    "args": {
      "Python parent id": 413, "Python module id": 2, "Python id": 435, "finished": true, "Ev Idx": 7265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564822118.523, "dur": 52.770,
    "args": {
      "Python parent id": 435, "Python id": 436, "finished": true, "Ev Idx": 7266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564822118.771, "dur": 0.533,
    "args": {
      "Python parent id": 436, "Python id": 437, "finished": true, "Ev Idx": 7267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564822120.318, "dur": 50.760,
    "args": {
      "Python parent id": 436, "Python id": 438, "finished": true, "Ev Idx": 7268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822120.602, "dur": 0.133,
    "args": {
      "Python parent id": 438, "Python id": 439, "finished": true, "Ev Idx": 7269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822120.971, "dur": 0.170,
    "args": {
      "Python parent id": 438, "Python id": 440, "finished": true, "Ev Idx": 7270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564822121.320, "dur": 49.686,
    "args": {
      "Python parent id": 438, "Python id": 441, "finished": true, "Ev Idx": 7271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564822172.297, "dur": 2.734,
    "args": {
      "Python parent id": 413, "Python id": 442, "finished": true, "Ev Idx": 7272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564822175.595, "dur": 2.973,
    "args": {
      "Python parent id": 413, "Python id": 443, "finished": true, "Ev Idx": 7273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564822181.282, "dur": 161.958,
    "args": {
      "Python parent id": 413, "Python id": 444, "finished": true, "Ev Idx": 7274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564822181.658, "dur": 3.779,
    "args": {
      "Python parent id": 444, "Python id": 445, "finished": true, "Ev Idx": 7275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564822185.992, "dur": 1.746,
    "args": {
      "Python parent id": 444, "Python id": 446, "finished": true, "Ev Idx": 7276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564822203.388, "dur": 54.384,
    "args": {
      "Python parent id": 444, "Python id": 447, "finished": true, "Ev Idx": 7277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564822234.468, "dur": 22.283,
    "args": {
      "Python parent id": 447, "Python id": 448, "finished": true, "Ev Idx": 7278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564822293.943, "dur": 29.974,
    "args": {
      "Python parent id": 444, "Python id": 449, "finished": true, "Ev Idx": 7279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564822311.292, "dur": 11.683,
    "args": {
      "Python parent id": 449, "Python id": 450, "finished": true, "Ev Idx": 7280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564822350.789, "dur": 8.259,
    "args": {
      "Python parent id": 413, "Python id": 451, "finished": true, "Ev Idx": 7281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564822352.459, "dur": 0.294,
    "args": {
      "Python parent id": 451, "Python id": 452, "finished": true, "Ev Idx": 7282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564822353.276, "dur": 0.071,
    "args": {
      "Python parent id": 451, "Python id": 453, "finished": true, "Ev Idx": 7283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564822357.188, "dur": 0.646,
    "args": {
      "Python parent id": 451, "Python id": 454, "finished": true, "Ev Idx": 7284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564822358.709, "dur": 0.074,
    "args": {
      "Python parent id": 451, "Python id": 455, "finished": true, "Ev Idx": 7285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564822361.174, "dur": 6.886,
    "args": {
      "Python parent id": 413, "Python id": 456, "finished": true, "Ev Idx": 7286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564822364.417, "dur": 3.512,
    "args": {
      "Python parent id": 456, "Python id": 457, "finished": true, "Ev Idx": 7287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564822364.822, "dur": 1.419,
    "args": {
      "Python parent id": 457, "Python id": 458, "finished": true, "Ev Idx": 7288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564822365.352, "dur": 0.751,
    "args": {
      "Python parent id": 458, "Python id": 459, "finished": true, "Ev Idx": 7289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564822366.528, "dur": 0.517,
    "args": {
      "Python parent id": 457, "Python id": 460, "finished": true, "Ev Idx": 7290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564822367.335, "dur": 0.503,
    "args": {
      "Python parent id": 457, "Python id": 461, "finished": true, "Ev Idx": 7291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564822373.453, "dur": 299.180,
    "args": {
      "Python parent id": 413, "Python id": 462, "finished": true, "Ev Idx": 7292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564822374.998, "dur": 52.356,
    "args": {
      "Python parent id": 462, "Python id": 463, "finished": true, "Ev Idx": 7293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564822395.359, "dur": 5.529,
    "args": {
      "Python parent id": 463, "Python id": 464, "finished": true, "Ev Idx": 7294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564822402.046, "dur": 25.170,
    "args": {
      "Python parent id": 463, "Python id": 465, "finished": true, "Ev Idx": 7295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564822428.606, "dur": 32.463,
    "args": {
      "Python parent id": 462, "Python id": 466, "finished": true, "Ev Idx": 7296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564822443.077, "dur": 3.133,
    "args": {
      "Python parent id": 466, "Python id": 467, "finished": true, "Ev Idx": 7297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564822446.764, "dur": 14.231,
    "args": {
      "Python parent id": 466, "Python id": 468, "finished": true, "Ev Idx": 7298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564822462.219, "dur": 2.898,
    "args": {
      "Python parent id": 462, "Python id": 469, "finished": true, "Ev Idx": 7299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564822466.768, "dur": 57.113,
    "args": {
      "Python parent id": 462, "Python id": 470, "finished": true, "Ev Idx": 7300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564822562.352, "dur": 22.494,
    "args": {
      "Python parent id": 462, "Python id": 471, "finished": true, "Ev Idx": 7301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564822563.679, "dur": 0.215,
    "args": {
      "Python parent id": 471, "Python id": 472, "finished": true, "Ev Idx": 7302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564822565.899, "dur": 18.789,
    "args": {
      "Python parent id": 471, "Python id": 473, "finished": true, "Ev Idx": 7303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564822585.841, "dur": 2.235,
    "args": {
      "Python parent id": 462, "Python id": 474, "finished": true, "Ev Idx": 7304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564822590.914, "dur": 12.075,
    "args": {
      "Python parent id": 462, "Python id": 475, "finished": true, "Ev Idx": 7305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564822591.119, "dur": 0.102,
    "args": {
      "Python parent id": 475, "Python id": 476, "finished": true, "Ev Idx": 7306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822596.935, "dur": 1.120,
    "args": {
      "Python parent id": 475, "Python id": 477, "finished": true, "Ev Idx": 7307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564822597.550, "dur": 0.390,
    "args": {
      "Python parent id": 477, "Python id": 478, "finished": true, "Ev Idx": 7308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564822599.490, "dur": 3.415,
    "args": {
      "Python parent id": 475, "Python id": 479, "finished": true, "Ev Idx": 7309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564822603.893, "dur": 46.916,
    "args": {
      "Python parent id": 462, "Python id": 480, "finished": true, "Ev Idx": 7310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564822651.914, "dur": 3.137,
    "args": {
      "Python parent id": 462, "Python id": 481, "finished": true, "Ev Idx": 7311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564822656.962, "dur": 14.607,
    "args": {
      "Python parent id": 462, "Python id": 482, "finished": true, "Ev Idx": 7312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564822676.593, "dur": 3.497,
    "args": {
      "Python parent id": 413, "Python id": 483, "finished": true, "Ev Idx": 7313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564822680.956, "dur": 0.267,
    "args": {
      "Python parent id": 413, "Python id": 484, "finished": true, "Ev Idx": 7314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822682.492, "dur": 0.859,
    "args": {
      "Python parent id": 413, "Python id": 485, "finished": true, "Ev Idx": 7315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_3", "pid": 537, "tid": 537,
    "ts": 7429564822687.755, "dur": 46.067,
    "args": {
      "Python parent id": 413, "Python module id": 3, "Python id": 486, "finished": true, "Ev Idx": 7316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564822688.643, "dur": 44.769,
    "args": {
      "Python parent id": 486, "Python id": 487, "finished": true, "Ev Idx": 7317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564822689.056, "dur": 0.975,
    "args": {
      "Python parent id": 487, "Python id": 488, "finished": true, "Ev Idx": 7318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564822691.310, "dur": 41.865,
    "args": {
      "Python parent id": 487, "Python id": 489, "finished": true, "Ev Idx": 7319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822691.713, "dur": 0.204,
    "args": {
      "Python parent id": 489, "Python id": 490, "finished": true, "Ev Idx": 7320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822692.190, "dur": 0.170,
    "args": {
      "Python parent id": 489, "Python id": 491, "finished": true, "Ev Idx": 7321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564822692.656, "dur": 40.399,
    "args": {
      "Python parent id": 489, "Python id": 492, "finished": true, "Ev Idx": 7322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822749.374, "dur": 0.673,
    "args": {
      "Python parent id": 397, "Python id": 493, "finished": true, "Ev Idx": 7323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_1", "pid": 537, "tid": 537,
    "ts": 7429564822752.946, "dur": 89.437,
    "args": {
      "Python parent id": 397, "Python module id": 1, "Python id": 494, "finished": true, "Ev Idx": 7324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564822753.591, "dur": 88.422,
    "args": {
      "Python parent id": 494, "Python id": 495, "finished": true, "Ev Idx": 7325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564822753.864, "dur": 0.486,
    "args": {
      "Python parent id": 495, "Python id": 496, "finished": true, "Ev Idx": 7326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564822755.446, "dur": 85.532,
    "args": {
      "Python parent id": 495, "Python id": 497, "finished": true, "Ev Idx": 7327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564822756.010, "dur": 1.826,
    "args": {
      "Python parent id": 497, "Python id": 498, "finished": true, "Ev Idx": 7328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564822758.270, "dur": 16.664,
    "args": {
      "Python parent id": 497, "Python id": 499, "finished": true, "Ev Idx": 7329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564822775.399, "dur": 17.345,
    "args": {
      "Python parent id": 497, "Python id": 500, "finished": true, "Ev Idx": 7330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564822805.204, "dur": 11.941,
    "args": {
      "Python parent id": 497, "Python id": 501, "finished": true, "Ev Idx": 7331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822829.633, "dur": 0.275,
    "args": {
      "Python parent id": 497, "Python id": 502, "finished": true, "Ev Idx": 7332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564822830.378, "dur": 1.424,
    "args": {
      "Python parent id": 497, "Python id": 503, "finished": true, "Ev Idx": 7333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822843.583, "dur": 0.343,
    "args": {
      "Python parent id": 397, "Python id": 504, "finished": true, "Ev Idx": 7334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_0", "pid": 537, "tid": 537,
    "ts": 7429564822847.611, "dur": 205.488,
    "args": {
      "Python parent id": 397, "Python module id": 0, "Python id": 505, "finished": true, "Ev Idx": 7335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564822848.300, "dur": 204.601,
    "args": {
      "Python parent id": 505, "Python id": 506, "finished": true, "Ev Idx": 7336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564822848.535, "dur": 0.426,
    "args": {
      "Python parent id": 506, "Python id": 507, "finished": true, "Ev Idx": 7337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564822850.979, "dur": 201.765,
    "args": {
      "Python parent id": 506, "Python id": 508, "finished": true, "Ev Idx": 7338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822851.587, "dur": 0.379,
    "args": {
      "Python parent id": 508, "Python id": 509, "finished": true, "Ev Idx": 7339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822852.387, "dur": 0.412,
    "args": {
      "Python parent id": 508, "Python id": 510, "finished": true, "Ev Idx": 7340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822853.045, "dur": 0.249,
    "args": {
      "Python parent id": 508, "Python id": 511, "finished": true, "Ev Idx": 7341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_4", "pid": 537, "tid": 537,
    "ts": 7429564822855.950, "dur": 56.457,
    "args": {
      "Python parent id": 508, "Python module id": 4, "Python id": 512, "finished": true, "Ev Idx": 7342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564822856.460, "dur": 55.672,
    "args": {
      "Python parent id": 512, "Python id": 513, "finished": true, "Ev Idx": 7343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564822856.625, "dur": 0.191,
    "args": {
      "Python parent id": 513, "Python id": 514, "finished": true, "Ev Idx": 7344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564822857.714, "dur": 54.307,
    "args": {
      "Python parent id": 513, "Python id": 515, "finished": true, "Ev Idx": 7345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822858.008, "dur": 0.184,
    "args": {
      "Python parent id": 515, "Python id": 516, "finished": true, "Ev Idx": 7346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822858.507, "dur": 0.165,
    "args": {
      "Python parent id": 515, "Python id": 517, "finished": true, "Ev Idx": 7347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564822858.803, "dur": 53.096,
    "args": {
      "Python parent id": 515, "Python id": 518, "finished": true, "Ev Idx": 7348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_0", "pid": 537, "tid": 537,
    "ts": 7429564822915.716, "dur": 23.355,
    "args": {
      "Python parent id": 508, "Python module id": 0, "Python id": 519, "finished": true, "Ev Idx": 7349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564822916.443, "dur": 22.401,
    "args": {
      "Python parent id": 519, "Python id": 520, "finished": true, "Ev Idx": 7350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564822916.655, "dur": 0.300,
    "args": {
      "Python parent id": 520, "Python id": 521, "finished": true, "Ev Idx": 7351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564822919.121, "dur": 19.644,
    "args": {
      "Python parent id": 520, "Python id": 522, "finished": true, "Ev Idx": 7352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564822920.953, "dur": 17.676,
    "args": {
      "Python parent id": 522, "Python id": 523, "finished": true, "Ev Idx": 7353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564822921.241, "dur": 1.023,
    "args": {
      "Python parent id": 523, "Python id": 524, "finished": true, "Ev Idx": 7354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564822924.408, "dur": 14.083,
    "args": {
      "Python parent id": 523, "Python id": 525, "finished": true, "Ev Idx": 7355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822939.804, "dur": 0.348,
    "args": {
      "Python parent id": 508, "Python id": 526, "finished": true, "Ev Idx": 7356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_5", "pid": 537, "tid": 537,
    "ts": 7429564822942.312, "dur": 44.660,
    "args": {
      "Python parent id": 508, "Python module id": 5, "Python id": 527, "finished": true, "Ev Idx": 7357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564822942.779, "dur": 43.967,
    "args": {
      "Python parent id": 527, "Python id": 528, "finished": true, "Ev Idx": 7358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564822942.944, "dur": 0.225,
    "args": {
      "Python parent id": 528, "Python id": 529, "finished": true, "Ev Idx": 7359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564822944.059, "dur": 42.556,
    "args": {
      "Python parent id": 528, "Python id": 530, "finished": true, "Ev Idx": 7360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822944.252, "dur": 0.151,
    "args": {
      "Python parent id": 530, "Python id": 531, "finished": true, "Ev Idx": 7361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564822944.593, "dur": 0.132,
    "args": {
      "Python parent id": 530, "Python id": 532, "finished": true, "Ev Idx": 7362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564822944.847, "dur": 41.656,
    "args": {
      "Python parent id": 530, "Python id": 533, "finished": true, "Ev Idx": 7363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_6", "pid": 537, "tid": 537,
    "ts": 7429564823002.269, "dur": 50.149,
    "args": {
      "Python parent id": 508, "Python module id": 6, "Python id": 534, "finished": true, "Ev Idx": 7364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564823002.950, "dur": 49.210,
    "args": {
      "Python parent id": 534, "Python id": 535, "finished": true, "Ev Idx": 7365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564823003.141, "dur": 0.362,
    "args": {
      "Python parent id": 535, "Python id": 536, "finished": true, "Ev Idx": 7366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564823004.126, "dur": 47.886,
    "args": {
      "Python parent id": 535, "Python id": 537, "finished": true, "Ev Idx": 7367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823004.381, "dur": 0.149,
    "args": {
      "Python parent id": 537, "Python id": 538, "finished": true, "Ev Idx": 7368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823004.738, "dur": 0.115,
    "args": {
      "Python parent id": 537, "Python id": 539, "finished": true, "Ev Idx": 7369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564823005.102, "dur": 46.848,
    "args": {
      "Python parent id": 537, "Python id": 540, "finished": true, "Ev Idx": 7370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_1", "pid": 537, "tid": 537,
    "ts": 7429564823072.577, "dur": 1271.402,
    "args": {
      "Python parent id": 64, "Python module id": 1, "Python id": 541, "finished": true, "Ev Idx": 7371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564823073.962, "dur": 1269.357,
    "args": {
      "Python parent id": 541, "Python id": 542, "finished": true, "Ev Idx": 7372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564823074.170, "dur": 0.353,
    "args": {
      "Python parent id": 542, "Python id": 543, "finished": true, "Ev Idx": 7373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564823076.395, "dur": 1265.789,
    "args": {
      "Python parent id": 542, "Python id": 544, "finished": true, "Ev Idx": 7374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823076.846, "dur": 0.467,
    "args": {
      "Python parent id": 544, "Python id": 545, "finished": true, "Ev Idx": 7375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_2", "pid": 537, "tid": 537,
    "ts": 7429564823079.285, "dur": 77.392,
    "args": {
      "Python parent id": 544, "Python module id": 2, "Python id": 546, "finished": true, "Ev Idx": 7376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564823079.718, "dur": 76.656,
    "args": {
      "Python parent id": 546, "Python id": 547, "finished": true, "Ev Idx": 7377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564823079.863, "dur": 0.148,
    "args": {
      "Python parent id": 547, "Python id": 548, "finished": true, "Ev Idx": 7378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564823080.837, "dur": 74.631,
    "args": {
      "Python parent id": 547, "Python id": 549, "finished": true, "Ev Idx": 7379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564823081.394, "dur": 1.662,
    "args": {
      "Python parent id": 549, "Python id": 550, "finished": true, "Ev Idx": 7380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564823083.376, "dur": 12.285,
    "args": {
      "Python parent id": 549, "Python id": 551, "finished": true, "Ev Idx": 7381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564823095.979, "dur": 14.242,
    "args": {
      "Python parent id": 549, "Python id": 552, "finished": true, "Ev Idx": 7382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564823122.010, "dur": 9.716,
    "args": {
      "Python parent id": 549, "Python id": 553, "finished": true, "Ev Idx": 7383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823143.910, "dur": 0.224,
    "args": {
      "Python parent id": 549, "Python id": 554, "finished": true, "Ev Idx": 7384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564823144.490, "dur": 1.129,
    "args": {
      "Python parent id": 549, "Python id": 555, "finished": true, "Ev Idx": 7385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823157.323, "dur": 0.345,
    "args": {
      "Python parent id": 544, "Python id": 556, "finished": true, "Ev Idx": 7386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_1", "pid": 537, "tid": 537,
    "ts": 7429564823160.641, "dur": 643.400,
    "args": {
      "Python parent id": 544, "Python module id": 1, "Python id": 557, "finished": true, "Ev Idx": 7387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564823161.858, "dur": 641.713,
    "args": {
      "Python parent id": 557, "Python id": 558, "finished": true, "Ev Idx": 7388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564823162.039, "dur": 0.315,
    "args": {
      "Python parent id": 558, "Python id": 559, "finished": true, "Ev Idx": 7389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564823163.777, "dur": 638.991,
    "args": {
      "Python parent id": 558, "Python id": 560, "finished": true, "Ev Idx": 7390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823166.574, "dur": 0.342,
    "args": {
      "Python parent id": 560, "Python id": 561, "finished": true, "Ev Idx": 7391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_7", "pid": 537, "tid": 537,
    "ts": 7429564823169.379, "dur": 42.481,
    "args": {
      "Python parent id": 560, "Python module id": 7, "Python id": 562, "finished": true, "Ev Idx": 7392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564823169.906, "dur": 41.685,
    "args": {
      "Python parent id": 562, "Python id": 563, "finished": true, "Ev Idx": 7393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564823170.079, "dur": 0.147,
    "args": {
      "Python parent id": 563, "Python id": 564, "finished": true, "Ev Idx": 7394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564823170.790, "dur": 40.725,
    "args": {
      "Python parent id": 563, "Python id": 565, "finished": true, "Ev Idx": 7395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823171.100, "dur": 0.134,
    "args": {
      "Python parent id": 565, "Python id": 566, "finished": true, "Ev Idx": 7396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823171.448, "dur": 0.114,
    "args": {
      "Python parent id": 565, "Python id": 567, "finished": true, "Ev Idx": 7397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564823171.682, "dur": 39.717,
    "args": {
      "Python parent id": 565, "Python id": 568, "finished": true, "Ev Idx": 7398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564823212.533, "dur": 3.236,
    "args": {
      "Python parent id": 560, "Python id": 569, "finished": true, "Ev Idx": 7399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564823216.222, "dur": 3.757,
    "args": {
      "Python parent id": 560, "Python id": 570, "finished": true, "Ev Idx": 7400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823220.824, "dur": 0.393,
    "args": {
      "Python parent id": 560, "Python id": 571, "finished": true, "Ev Idx": 7401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_8", "pid": 537, "tid": 537,
    "ts": 7429564823223.457, "dur": 44.967,
    "args": {
      "Python parent id": 560, "Python module id": 8, "Python id": 572, "finished": true, "Ev Idx": 7402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564823224.164, "dur": 43.996,
    "args": {
      "Python parent id": 572, "Python id": 573, "finished": true, "Ev Idx": 7403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564823224.391, "dur": 0.239,
    "args": {
      "Python parent id": 573, "Python id": 574, "finished": true, "Ev Idx": 7404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564823225.291, "dur": 42.781,
    "args": {
      "Python parent id": 573, "Python id": 575, "finished": true, "Ev Idx": 7405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823225.577, "dur": 0.164,
    "args": {
      "Python parent id": 575, "Python id": 576, "finished": true, "Ev Idx": 7406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823225.949, "dur": 0.130,
    "args": {
      "Python parent id": 575, "Python id": 577, "finished": true, "Ev Idx": 7407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564823226.322, "dur": 41.701,
    "args": {
      "Python parent id": 575, "Python id": 578, "finished": true, "Ev Idx": 7408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564823268.936, "dur": 2.601,
    "args": {
      "Python parent id": 560, "Python id": 579, "finished": true, "Ev Idx": 7409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564823272.009, "dur": 3.236,
    "args": {
      "Python parent id": 560, "Python id": 580, "finished": true, "Ev Idx": 7410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823275.990, "dur": 0.436,
    "args": {
      "Python parent id": 560, "Python id": 581, "finished": true, "Ev Idx": 7411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_9", "pid": 537, "tid": 537,
    "ts": 7429564823278.816, "dur": 42.371,
    "args": {
      "Python parent id": 560, "Python module id": 9, "Python id": 582, "finished": true, "Ev Idx": 7412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564823279.353, "dur": 41.547,
    "args": {
      "Python parent id": 582, "Python id": 583, "finished": true, "Ev Idx": 7413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564823279.543, "dur": 0.252,
    "args": {
      "Python parent id": 583, "Python id": 584, "finished": true, "Ev Idx": 7414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564823280.354, "dur": 40.370,
    "args": {
      "Python parent id": 583, "Python id": 585, "finished": true, "Ev Idx": 7415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823280.566, "dur": 0.122,
    "args": {
      "Python parent id": 585, "Python id": 586, "finished": true, "Ev Idx": 7416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823280.863, "dur": 0.114,
    "args": {
      "Python parent id": 585, "Python id": 587, "finished": true, "Ev Idx": 7417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564823281.085, "dur": 39.587,
    "args": {
      "Python parent id": 585, "Python id": 588, "finished": true, "Ev Idx": 7418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564823321.683, "dur": 2.618,
    "args": {
      "Python parent id": 560, "Python id": 589, "finished": true, "Ev Idx": 7419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564823324.831, "dur": 2.755,
    "args": {
      "Python parent id": 560, "Python id": 590, "finished": true, "Ev Idx": 7420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564823328.686, "dur": 143.242,
    "args": {
      "Python parent id": 560, "Python id": 591, "finished": true, "Ev Idx": 7421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564823328.944, "dur": 3.513,
    "args": {
      "Python parent id": 591, "Python id": 592, "finished": true, "Ev Idx": 7422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564823332.780, "dur": 1.842,
    "args": {
      "Python parent id": 591, "Python id": 593, "finished": true, "Ev Idx": 7423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564823346.495, "dur": 43.374,
    "args": {
      "Python parent id": 591, "Python id": 594, "finished": true, "Ev Idx": 7424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564823371.312, "dur": 17.844,
    "args": {
      "Python parent id": 594, "Python id": 595, "finished": true, "Ev Idx": 7425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564823422.498, "dur": 30.739,
    "args": {
      "Python parent id": 591, "Python id": 596, "finished": true, "Ev Idx": 7426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564823439.743, "dur": 12.553,
    "args": {
      "Python parent id": 596, "Python id": 597, "finished": true, "Ev Idx": 7427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564823476.468, "dur": 3.585,
    "args": {
      "Python parent id": 560, "Python id": 598, "finished": true, "Ev Idx": 7428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564823477.141, "dur": 0.203,
    "args": {
      "Python parent id": 598, "Python id": 599, "finished": true, "Ev Idx": 7429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564823477.623, "dur": 0.144,
    "args": {
      "Python parent id": 598, "Python id": 600, "finished": true, "Ev Idx": 7430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564823478.989, "dur": 0.213,
    "args": {
      "Python parent id": 598, "Python id": 601, "finished": true, "Ev Idx": 7431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564823479.692, "dur": 0.122,
    "args": {
      "Python parent id": 598, "Python id": 602, "finished": true, "Ev Idx": 7432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564823481.211, "dur": 5.900,
    "args": {
      "Python parent id": 560, "Python id": 603, "finished": true, "Ev Idx": 7433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564823483.629, "dur": 3.278,
    "args": {
      "Python parent id": 603, "Python id": 604, "finished": true, "Ev Idx": 7434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564823483.830, "dur": 1.418,
    "args": {
      "Python parent id": 604, "Python id": 605, "finished": true, "Ev Idx": 7435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564823484.410, "dur": 0.683,
    "args": {
      "Python parent id": 605, "Python id": 606, "finished": true, "Ev Idx": 7436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564823485.471, "dur": 0.583,
    "args": {
      "Python parent id": 604, "Python id": 607, "finished": true, "Ev Idx": 7437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564823486.329, "dur": 0.507,
    "args": {
      "Python parent id": 604, "Python id": 608, "finished": true, "Ev Idx": 7438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564823489.661, "dur": 255.459,
    "args": {
      "Python parent id": 560, "Python id": 609, "finished": true, "Ev Idx": 7439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564823490.148, "dur": 43.994,
    "args": {
      "Python parent id": 609, "Python id": 610, "finished": true, "Ev Idx": 7440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564823507.575, "dur": 4.996,
    "args": {
      "Python parent id": 610, "Python id": 611, "finished": true, "Ev Idx": 7441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564823513.374, "dur": 20.638,
    "args": {
      "Python parent id": 610, "Python id": 612, "finished": true, "Ev Idx": 7442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564823534.971, "dur": 32.491,
    "args": {
      "Python parent id": 609, "Python id": 613, "finished": true, "Ev Idx": 7443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564823549.153, "dur": 2.774,
    "args": {
      "Python parent id": 613, "Python id": 614, "finished": true, "Ev Idx": 7444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564823552.377, "dur": 14.968,
    "args": {
      "Python parent id": 613, "Python id": 615, "finished": true, "Ev Idx": 7445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564823568.288, "dur": 3.274,
    "args": {
      "Python parent id": 609, "Python id": 616, "finished": true, "Ev Idx": 7446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564823571.790, "dur": 46.684,
    "args": {
      "Python parent id": 609, "Python id": 617, "finished": true, "Ev Idx": 7447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564823652.508, "dur": 16.490,
    "args": {
      "Python parent id": 609, "Python id": 618, "finished": true, "Ev Idx": 7448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564823652.680, "dur": 0.152,
    "args": {
      "Python parent id": 618, "Python id": 619, "finished": true, "Ev Idx": 7449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564823653.642, "dur": 15.216,
    "args": {
      "Python parent id": 618, "Python id": 620, "finished": true, "Ev Idx": 7450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564823669.881, "dur": 1.569,
    "args": {
      "Python parent id": 609, "Python id": 621, "finished": true, "Ev Idx": 7451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564823673.174, "dur": 9.734,
    "args": {
      "Python parent id": 609, "Python id": 622, "finished": true, "Ev Idx": 7452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564823673.258, "dur": 0.100,
    "args": {
      "Python parent id": 622, "Python id": 623, "finished": true, "Ev Idx": 7453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823678.031, "dur": 1.632,
    "args": {
      "Python parent id": 622, "Python id": 624, "finished": true, "Ev Idx": 7454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564823679.197, "dur": 0.388,
    "args": {
      "Python parent id": 624, "Python id": 625, "finished": true, "Ev Idx": 7455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564823680.188, "dur": 2.644,
    "args": {
      "Python parent id": 622, "Python id": 626, "finished": true, "Ev Idx": 7456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564823683.682, "dur": 40.504,
    "args": {
      "Python parent id": 609, "Python id": 627, "finished": true, "Ev Idx": 7457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564823724.813, "dur": 3.197,
    "args": {
      "Python parent id": 609, "Python id": 628, "finished": true, "Ev Idx": 7458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564823728.724, "dur": 15.387,
    "args": {
      "Python parent id": 609, "Python id": 629, "finished": true, "Ev Idx": 7459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564823747.978, "dur": 3.172,
    "args": {
      "Python parent id": 560, "Python id": 630, "finished": true, "Ev Idx": 7460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564823751.924, "dur": 0.277,
    "args": {
      "Python parent id": 560, "Python id": 631, "finished": true, "Ev Idx": 7461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823753.207, "dur": 0.897,
    "args": {
      "Python parent id": 560, "Python id": 632, "finished": true, "Ev Idx": 7462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_10", "pid": 537, "tid": 537,
    "ts": 7429564823758.191, "dur": 43.972,
    "args": {
      "Python parent id": 560, "Python module id": 10, "Python id": 633, "finished": true, "Ev Idx": 7463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564823759.213, "dur": 42.618,
    "args": {
      "Python parent id": 633, "Python id": 634, "finished": true, "Ev Idx": 7464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564823759.657, "dur": 0.957,
    "args": {
      "Python parent id": 634, "Python id": 635, "finished": true, "Ev Idx": 7465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564823761.705, "dur": 39.902,
    "args": {
      "Python parent id": 634, "Python id": 636, "finished": true, "Ev Idx": 7466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823762.090, "dur": 0.506,
    "args": {
      "Python parent id": 636, "Python id": 637, "finished": true, "Ev Idx": 7467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823762.858, "dur": 0.155,
    "args": {
      "Python parent id": 636, "Python id": 638, "finished": true, "Ev Idx": 7468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564823763.153, "dur": 38.299,
    "args": {
      "Python parent id": 636, "Python id": 639, "finished": true, "Ev Idx": 7469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564823900.104, "dur": 0.738,
    "args": {
      "Python parent id": 544, "Python id": 640, "finished": true, "Ev Idx": 7470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_3", "pid": 537, "tid": 537,
    "ts": 7429564823904.332, "dur": 114.760,
    "args": {
      "Python parent id": 544, "Python module id": 3, "Python id": 641, "finished": true, "Ev Idx": 7471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564823905.104, "dur": 113.629,
    "args": {
      "Python parent id": 641, "Python id": 642, "finished": true, "Ev Idx": 7472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564823905.446, "dur": 0.649,
    "args": {
      "Python parent id": 642, "Python id": 643, "finished": true, "Ev Idx": 7473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564823907.849, "dur": 109.761,
    "args": {
      "Python parent id": 642, "Python id": 644, "finished": true, "Ev Idx": 7474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564823908.492, "dur": 2.444,
    "args": {
      "Python parent id": 644, "Python id": 645, "finished": true, "Ev Idx": 7475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564823911.275, "dur": 16.606,
    "args": {
      "Python parent id": 644, "Python id": 646, "finished": true, "Ev Idx": 7476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564823928.278, "dur": 18.374,
    "args": {
      "Python parent id": 644, "Python id": 647, "finished": true, "Ev Idx": 7477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564823980.738, "dur": 13.931,
    "args": {
      "Python parent id": 644, "Python id": 648, "finished": true, "Ev Idx": 7478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824007.055, "dur": 0.206,
    "args": {
      "Python parent id": 644, "Python id": 649, "finished": true, "Ev Idx": 7479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564824007.713, "dur": 1.366,
    "args": {
      "Python parent id": 644, "Python id": 650, "finished": true, "Ev Idx": 7480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824020.019, "dur": 0.461,
    "args": {
      "Python parent id": 544, "Python id": 651, "finished": true, "Ev Idx": 7481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_1", "pid": 537, "tid": 537,
    "ts": 7429564824022.385, "dur": 308.221,
    "args": {
      "Python parent id": 544, "Python module id": 1, "Python id": 652, "finished": true, "Ev Idx": 7482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824022.885, "dur": 307.392,
    "args": {
      "Python parent id": 652, "Python id": 653, "finished": true, "Ev Idx": 7483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824023.122, "dur": 0.376,
    "args": {
      "Python parent id": 653, "Python id": 654, "finished": true, "Ev Idx": 7484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564824024.650, "dur": 305.420,
    "args": {
      "Python parent id": 653, "Python id": 655, "finished": true, "Ev Idx": 7485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824024.999, "dur": 0.510,
    "args": {
      "Python parent id": 655, "Python id": 656, "finished": true, "Ev Idx": 7486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824025.906, "dur": 0.303,
    "args": {
      "Python parent id": 655, "Python id": 657, "finished": true, "Ev Idx": 7487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824026.481, "dur": 0.243,
    "args": {
      "Python parent id": 655, "Python id": 658, "finished": true, "Ev Idx": 7488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_11", "pid": 537, "tid": 537,
    "ts": 7429564824029.516, "dur": 70.766,
    "args": {
      "Python parent id": 655, "Python module id": 11, "Python id": 659, "finished": true, "Ev Idx": 7489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824030.169, "dur": 69.761,
    "args": {
      "Python parent id": 659, "Python id": 660, "finished": true, "Ev Idx": 7490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824030.311, "dur": 0.167,
    "args": {
      "Python parent id": 660, "Python id": 661, "finished": true, "Ev Idx": 7491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564824031.063, "dur": 68.512,
    "args": {
      "Python parent id": 660, "Python id": 662, "finished": true, "Ev Idx": 7492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824031.397, "dur": 0.161,
    "args": {
      "Python parent id": 662, "Python id": 663, "finished": true, "Ev Idx": 7493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824031.812, "dur": 0.151,
    "args": {
      "Python parent id": 662, "Python id": 664, "finished": true, "Ev Idx": 7494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564824032.118, "dur": 67.355,
    "args": {
      "Python parent id": 662, "Python id": 665, "finished": true, "Ev Idx": 7495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_1", "pid": 537, "tid": 537,
    "ts": 7429564824102.879, "dur": 56.235,
    "args": {
      "Python parent id": 655, "Python module id": 1, "Python id": 666, "finished": true, "Ev Idx": 7496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824103.674, "dur": 55.206,
    "args": {
      "Python parent id": 666, "Python id": 667, "finished": true, "Ev Idx": 7497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824103.977, "dur": 0.448,
    "args": {
      "Python parent id": 667, "Python id": 668, "finished": true, "Ev Idx": 7498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564824105.739, "dur": 52.992,
    "args": {
      "Python parent id": 667, "Python id": 669, "finished": true, "Ev Idx": 7499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564824106.828, "dur": 51.771,
    "args": {
      "Python parent id": 669, "Python id": 670, "finished": true, "Ev Idx": 7500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564824107.058, "dur": 0.455,
    "args": {
      "Python parent id": 670, "Python id": 671, "finished": true, "Ev Idx": 7501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564824108.527, "dur": 49.988,
    "args": {
      "Python parent id": 670, "Python id": 672, "finished": true, "Ev Idx": 7502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824159.919, "dur": 0.443,
    "args": {
      "Python parent id": 655, "Python id": 673, "finished": true, "Ev Idx": 7503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_12", "pid": 537, "tid": 537,
    "ts": 7429564824162.707, "dur": 106.168,
    "args": {
      "Python parent id": 655, "Python module id": 12, "Python id": 674, "finished": true, "Ev Idx": 7504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824163.404, "dur": 105.180,
    "args": {
      "Python parent id": 674, "Python id": 675, "finished": true, "Ev Idx": 7505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824163.618, "dur": 0.195,
    "args": {
      "Python parent id": 675, "Python id": 676, "finished": true, "Ev Idx": 7506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564824164.507, "dur": 103.895,
    "args": {
      "Python parent id": 675, "Python id": 677, "finished": true, "Ev Idx": 7507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824164.799, "dur": 0.164,
    "args": {
      "Python parent id": 677, "Python id": 678, "finished": true, "Ev Idx": 7508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824165.153, "dur": 0.159,
    "args": {
      "Python parent id": 677, "Python id": 679, "finished": true, "Ev Idx": 7509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564824165.451, "dur": 102.785,
    "args": {
      "Python parent id": 677, "Python id": 680, "finished": true, "Ev Idx": 7510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_13", "pid": 537, "tid": 537,
    "ts": 7429564824283.231, "dur": 46.487,
    "args": {
      "Python parent id": 655, "Python module id": 13, "Python id": 681, "finished": true, "Ev Idx": 7511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824284.119, "dur": 45.420,
    "args": {
      "Python parent id": 681, "Python id": 682, "finished": true, "Ev Idx": 7512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824284.412, "dur": 0.477,
    "args": {
      "Python parent id": 682, "Python id": 683, "finished": true, "Ev Idx": 7513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564824285.608, "dur": 43.821,
    "args": {
      "Python parent id": 682, "Python id": 684, "finished": true, "Ev Idx": 7514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824285.950, "dur": 0.223,
    "args": {
      "Python parent id": 684, "Python id": 685, "finished": true, "Ev Idx": 7515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824286.400, "dur": 0.178,
    "args": {
      "Python parent id": 684, "Python id": 686, "finished": true, "Ev Idx": 7516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564824286.693, "dur": 42.657,
    "args": {
      "Python parent id": 684, "Python id": 687, "finished": true, "Ev Idx": 7517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_2", "pid": 537, "tid": 537,
    "ts": 7429564824351.514, "dur": 1233.408,
    "args": {
      "Python parent id": 64, "Python module id": 2, "Python id": 688, "finished": true, "Ev Idx": 7518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824353.004, "dur": 1231.049,
    "args": {
      "Python parent id": 688, "Python id": 689, "finished": true, "Ev Idx": 7519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824353.241, "dur": 0.358,
    "args": {
      "Python parent id": 689, "Python id": 690, "finished": true, "Ev Idx": 7520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564824355.353, "dur": 1227.760,
    "args": {
      "Python parent id": 689, "Python id": 691, "finished": true, "Ev Idx": 7521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824355.905, "dur": 0.704,
    "args": {
      "Python parent id": 691, "Python id": 692, "finished": true, "Ev Idx": 7522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_4", "pid": 537, "tid": 537,
    "ts": 7429564824358.773, "dur": 122.001,
    "args": {
      "Python parent id": 691, "Python module id": 4, "Python id": 693, "finished": true, "Ev Idx": 7523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824359.310, "dur": 121.103,
    "args": {
      "Python parent id": 693, "Python id": 694, "finished": true, "Ev Idx": 7524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824359.473, "dur": 0.169,
    "args": {
      "Python parent id": 694, "Python id": 695, "finished": true, "Ev Idx": 7525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564824360.421, "dur": 118.803,
    "args": {
      "Python parent id": 694, "Python id": 696, "finished": true, "Ev Idx": 7526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564824361.054, "dur": 2.341,
    "args": {
      "Python parent id": 696, "Python id": 697, "finished": true, "Ev Idx": 7527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564824363.698, "dur": 12.440,
    "args": {
      "Python parent id": 696, "Python id": 698, "finished": true, "Ev Idx": 7528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564824376.545, "dur": 14.755,
    "args": {
      "Python parent id": 696, "Python id": 699, "finished": true, "Ev Idx": 7529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564824404.017, "dur": 10.042,
    "args": {
      "Python parent id": 696, "Python id": 700, "finished": true, "Ev Idx": 7530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824424.585, "dur": 0.257,
    "args": {
      "Python parent id": 696, "Python id": 701, "finished": true, "Ev Idx": 7531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564824425.346, "dur": 1.278,
    "args": {
      "Python parent id": 696, "Python id": 702, "finished": true, "Ev Idx": 7532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824481.565, "dur": 0.379,
    "args": {
      "Python parent id": 691, "Python id": 703, "finished": true, "Ev Idx": 7533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_2", "pid": 537, "tid": 537,
    "ts": 7429564824490.175, "dur": 675.492,
    "args": {
      "Python parent id": 691, "Python module id": 2, "Python id": 704, "finished": true, "Ev Idx": 7534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824491.786, "dur": 673.464,
    "args": {
      "Python parent id": 704, "Python id": 705, "finished": true, "Ev Idx": 7535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824492.088, "dur": 0.498,
    "args": {
      "Python parent id": 705, "Python id": 706, "finished": true, "Ev Idx": 7536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564824494.811, "dur": 669.679,
    "args": {
      "Python parent id": 705, "Python id": 707, "finished": true, "Ev Idx": 7537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824498.774, "dur": 0.436,
    "args": {
      "Python parent id": 707, "Python id": 708, "finished": true, "Ev Idx": 7538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_14", "pid": 537, "tid": 537,
    "ts": 7429564824501.976, "dur": 45.619,
    "args": {
      "Python parent id": 707, "Python module id": 14, "Python id": 709, "finished": true, "Ev Idx": 7539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824502.653, "dur": 44.712,
    "args": {
      "Python parent id": 709, "Python id": 710, "finished": true, "Ev Idx": 7540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824502.804, "dur": 0.182,
    "args": {
      "Python parent id": 710, "Python id": 711, "finished": true, "Ev Idx": 7541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564824503.800, "dur": 43.429,
    "args": {
      "Python parent id": 710, "Python id": 712, "finished": true, "Ev Idx": 7542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824504.331, "dur": 0.134,
    "args": {
      "Python parent id": 712, "Python id": 713, "finished": true, "Ev Idx": 7543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824504.697, "dur": 0.139,
    "args": {
      "Python parent id": 712, "Python id": 714, "finished": true, "Ev Idx": 7544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564824504.960, "dur": 42.191,
    "args": {
      "Python parent id": 712, "Python id": 715, "finished": true, "Ev Idx": 7545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564824548.412, "dur": 4.076,
    "args": {
      "Python parent id": 707, "Python id": 716, "finished": true, "Ev Idx": 7546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564824552.971, "dur": 4.232,
    "args": {
      "Python parent id": 707, "Python id": 717, "finished": true, "Ev Idx": 7547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824558.213, "dur": 0.372,
    "args": {
      "Python parent id": 707, "Python id": 718, "finished": true, "Ev Idx": 7548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_15", "pid": 537, "tid": 537,
    "ts": 7429564824561.424, "dur": 45.782,
    "args": {
      "Python parent id": 707, "Python module id": 15, "Python id": 719, "finished": true, "Ev Idx": 7549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824562.153, "dur": 44.701,
    "args": {
      "Python parent id": 719, "Python id": 720, "finished": true, "Ev Idx": 7550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824562.391, "dur": 0.283,
    "args": {
      "Python parent id": 720, "Python id": 721, "finished": true, "Ev Idx": 7551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564824563.428, "dur": 43.328,
    "args": {
      "Python parent id": 720, "Python id": 722, "finished": true, "Ev Idx": 7552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824563.631, "dur": 0.126,
    "args": {
      "Python parent id": 722, "Python id": 723, "finished": true, "Ev Idx": 7553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824563.974, "dur": 0.237,
    "args": {
      "Python parent id": 722, "Python id": 724, "finished": true, "Ev Idx": 7554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564824564.332, "dur": 42.370,
    "args": {
      "Python parent id": 722, "Python id": 725, "finished": true, "Ev Idx": 7555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564824607.768, "dur": 2.707,
    "args": {
      "Python parent id": 707, "Python id": 726, "finished": true, "Ev Idx": 7556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564824611.142, "dur": 2.598,
    "args": {
      "Python parent id": 707, "Python id": 727, "finished": true, "Ev Idx": 7557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824614.460, "dur": 0.522,
    "args": {
      "Python parent id": 707, "Python id": 728, "finished": true, "Ev Idx": 7558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_16", "pid": 537, "tid": 537,
    "ts": 7429564824617.730, "dur": 40.436,
    "args": {
      "Python parent id": 707, "Python module id": 16, "Python id": 729, "finished": true, "Ev Idx": 7559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564824618.303, "dur": 39.622,
    "args": {
      "Python parent id": 729, "Python id": 730, "finished": true, "Ev Idx": 7560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564824618.537, "dur": 0.264,
    "args": {
      "Python parent id": 730, "Python id": 731, "finished": true, "Ev Idx": 7561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564824619.694, "dur": 38.085,
    "args": {
      "Python parent id": 730, "Python id": 732, "finished": true, "Ev Idx": 7562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824619.949, "dur": 0.178,
    "args": {
      "Python parent id": 732, "Python id": 733, "finished": true, "Ev Idx": 7563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564824620.327, "dur": 0.145,
    "args": {
      "Python parent id": 732, "Python id": 734, "finished": true, "Ev Idx": 7564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564824620.573, "dur": 37.158,
    "args": {
      "Python parent id": 732, "Python id": 735, "finished": true, "Ev Idx": 7565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564824658.713, "dur": 2.610,
    "args": {
      "Python parent id": 707, "Python id": 736, "finished": true, "Ev Idx": 7566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564824661.801, "dur": 2.922,
    "args": {
      "Python parent id": 707, "Python id": 737, "finished": true, "Ev Idx": 7567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564824666.237, "dur": 155.366,
    "args": {
      "Python parent id": 707, "Python id": 738, "finished": true, "Ev Idx": 7568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564824666.612, "dur": 4.052,
    "args": {
      "Python parent id": 738, "Python id": 739, "finished": true, "Ev Idx": 7569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564824670.988, "dur": 1.827,
    "args": {
      "Python parent id": 738, "Python id": 740, "finished": true, "Ev Idx": 7570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564824684.965, "dur": 56.323,
    "args": {
      "Python parent id": 738, "Python id": 741, "finished": true, "Ev Idx": 7571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564824716.626, "dur": 23.857,
    "args": {
      "Python parent id": 741, "Python id": 742, "finished": true, "Ev Idx": 7572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564824772.399, "dur": 32.365,
    "args": {
      "Python parent id": 738, "Python id": 743, "finished": true, "Ev Idx": 7573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564824790.499, "dur": 13.448,
    "args": {
      "Python parent id": 743, "Python id": 744, "finished": true, "Ev Idx": 7574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564824826.341, "dur": 4.586,
    "args": {
      "Python parent id": 707, "Python id": 745, "finished": true, "Ev Idx": 7575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564824827.207, "dur": 0.457,
    "args": {
      "Python parent id": 745, "Python id": 746, "finished": true, "Ev Idx": 7576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564824828.063, "dur": 0.070,
    "args": {
      "Python parent id": 745, "Python id": 747, "finished": true, "Ev Idx": 7577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564824829.490, "dur": 0.367,
    "args": {
      "Python parent id": 745, "Python id": 748, "finished": true, "Ev Idx": 7578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564824830.394, "dur": 0.123,
    "args": {
      "Python parent id": 745, "Python id": 749, "finished": true, "Ev Idx": 7579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564824832.344, "dur": 6.808,
    "args": {
      "Python parent id": 707, "Python id": 750, "finished": true, "Ev Idx": 7580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564824835.703, "dur": 3.307,
    "args": {
      "Python parent id": 750, "Python id": 751, "finished": true, "Ev Idx": 7581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564824835.939, "dur": 1.552,
    "args": {
      "Python parent id": 751, "Python id": 752, "finished": true, "Ev Idx": 7582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564824836.513, "dur": 0.809,
    "args": {
      "Python parent id": 752, "Python id": 753, "finished": true, "Ev Idx": 7583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564824837.792, "dur": 0.451,
    "args": {
      "Python parent id": 751, "Python id": 754, "finished": true, "Ev Idx": 7584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564824838.528, "dur": 0.407,
    "args": {
      "Python parent id": 751, "Python id": 755, "finished": true, "Ev Idx": 7585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564824842.006, "dur": 267.186,
    "args": {
      "Python parent id": 707, "Python id": 756, "finished": true, "Ev Idx": 7586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564824842.648, "dur": 49.837,
    "args": {
      "Python parent id": 756, "Python id": 757, "finished": true, "Ev Idx": 7587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564824861.731, "dur": 5.602,
    "args": {
      "Python parent id": 757, "Python id": 758, "finished": true, "Ev Idx": 7588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564824868.257, "dur": 24.086,
    "args": {
      "Python parent id": 757, "Python id": 759, "finished": true, "Ev Idx": 7589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564824893.435, "dur": 33.892,
    "args": {
      "Python parent id": 756, "Python id": 760, "finished": true, "Ev Idx": 7590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564824911.645, "dur": 3.001,
    "args": {
      "Python parent id": 760, "Python id": 761, "finished": true, "Ev Idx": 7591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564824915.158, "dur": 12.079,
    "args": {
      "Python parent id": 760, "Python id": 762, "finished": true, "Ev Idx": 7592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564824928.292, "dur": 3.017,
    "args": {
      "Python parent id": 756, "Python id": 763, "finished": true, "Ev Idx": 7593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564824931.604, "dur": 46.686,
    "args": {
      "Python parent id": 756, "Python id": 764, "finished": true, "Ev Idx": 7594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564825011.610, "dur": 18.121,
    "args": {
      "Python parent id": 756, "Python id": 765, "finished": true, "Ev Idx": 7595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564825011.828, "dur": 0.408,
    "args": {
      "Python parent id": 765, "Python id": 766, "finished": true, "Ev Idx": 7596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564825013.127, "dur": 16.459,
    "args": {
      "Python parent id": 765, "Python id": 767, "finished": true, "Ev Idx": 7597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564825030.727, "dur": 1.778,
    "args": {
      "Python parent id": 756, "Python id": 768, "finished": true, "Ev Idx": 7598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564825034.148, "dur": 11.410,
    "args": {
      "Python parent id": 756, "Python id": 769, "finished": true, "Ev Idx": 7599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564825034.452, "dur": 0.106,
    "args": {
      "Python parent id": 769, "Python id": 770, "finished": true, "Ev Idx": 7600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825040.396, "dur": 1.761,
    "args": {
      "Python parent id": 769, "Python id": 771, "finished": true, "Ev Idx": 7601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564825041.570, "dur": 0.502,
    "args": {
      "Python parent id": 771, "Python id": 772, "finished": true, "Ev Idx": 7602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564825042.559, "dur": 2.914,
    "args": {
      "Python parent id": 769, "Python id": 773, "finished": true, "Ev Idx": 7603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564825046.253, "dur": 43.836,
    "args": {
      "Python parent id": 756, "Python id": 774, "finished": true, "Ev Idx": 7604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564825090.676, "dur": 2.986,
    "args": {
      "Python parent id": 756, "Python id": 775, "finished": true, "Ev Idx": 7605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564825094.402, "dur": 13.695,
    "args": {
      "Python parent id": 756, "Python id": 776, "finished": true, "Ev Idx": 7606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564825112.075, "dur": 3.452,
    "args": {
      "Python parent id": 707, "Python id": 777, "finished": true, "Ev Idx": 7607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564825116.210, "dur": 0.263,
    "args": {
      "Python parent id": 707, "Python id": 778, "finished": true, "Ev Idx": 7608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825117.377, "dur": 0.765,
    "args": {
      "Python parent id": 707, "Python id": 779, "finished": true, "Ev Idx": 7609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_17", "pid": 537, "tid": 537,
    "ts": 7429564825122.106, "dur": 41.861,
    "args": {
      "Python parent id": 707, "Python module id": 17, "Python id": 780, "finished": true, "Ev Idx": 7610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825122.950, "dur": 40.628,
    "args": {
      "Python parent id": 780, "Python id": 781, "finished": true, "Ev Idx": 7611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825123.421, "dur": 0.926,
    "args": {
      "Python parent id": 781, "Python id": 782, "finished": true, "Ev Idx": 7612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564825125.389, "dur": 37.962,
    "args": {
      "Python parent id": 781, "Python id": 783, "finished": true, "Ev Idx": 7613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825125.688, "dur": 0.388,
    "args": {
      "Python parent id": 783, "Python id": 784, "finished": true, "Ev Idx": 7614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825126.346, "dur": 0.161,
    "args": {
      "Python parent id": 783, "Python id": 785, "finished": true, "Ev Idx": 7615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564825126.665, "dur": 36.593,
    "args": {
      "Python parent id": 783, "Python id": 786, "finished": true, "Ev Idx": 7616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825177.725, "dur": 0.565,
    "args": {
      "Python parent id": 691, "Python id": 787, "finished": true, "Ev Idx": 7617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_5", "pid": 537, "tid": 537,
    "ts": 7429564825181.108, "dur": 81.951,
    "args": {
      "Python parent id": 691, "Python module id": 5, "Python id": 788, "finished": true, "Ev Idx": 7618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825181.683, "dur": 81.045,
    "args": {
      "Python parent id": 788, "Python id": 789, "finished": true, "Ev Idx": 7619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825181.866, "dur": 0.363,
    "args": {
      "Python parent id": 789, "Python id": 790, "finished": true, "Ev Idx": 7620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564825183.117, "dur": 78.681,
    "args": {
      "Python parent id": 789, "Python id": 791, "finished": true, "Ev Idx": 7621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564825183.628, "dur": 1.688,
    "args": {
      "Python parent id": 791, "Python id": 792, "finished": true, "Ev Idx": 7622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564825185.597, "dur": 13.596,
    "args": {
      "Python parent id": 791, "Python id": 793, "finished": true, "Ev Idx": 7623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564825199.566, "dur": 15.080,
    "args": {
      "Python parent id": 791, "Python id": 794, "finished": true, "Ev Idx": 7624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564825227.289, "dur": 11.607,
    "args": {
      "Python parent id": 791, "Python id": 795, "finished": true, "Ev Idx": 7625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825250.427, "dur": 0.217,
    "args": {
      "Python parent id": 791, "Python id": 796, "finished": true, "Ev Idx": 7626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564825251.035, "dur": 1.539,
    "args": {
      "Python parent id": 791, "Python id": 797, "finished": true, "Ev Idx": 7627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825263.767, "dur": 0.493,
    "args": {
      "Python parent id": 691, "Python id": 798, "finished": true, "Ev Idx": 7628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_2", "pid": 537, "tid": 537,
    "ts": 7429564825265.945, "dur": 304.213,
    "args": {
      "Python parent id": 691, "Python module id": 2, "Python id": 799, "finished": true, "Ev Idx": 7629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825266.364, "dur": 303.420,
    "args": {
      "Python parent id": 799, "Python id": 800, "finished": true, "Ev Idx": 7630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825266.660, "dur": 0.353,
    "args": {
      "Python parent id": 800, "Python id": 801, "finished": true, "Ev Idx": 7631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564825268.193, "dur": 301.492,
    "args": {
      "Python parent id": 800, "Python id": 802, "finished": true, "Ev Idx": 7632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825268.462, "dur": 0.353,
    "args": {
      "Python parent id": 802, "Python id": 803, "finished": true, "Ev Idx": 7633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825269.193, "dur": 0.314,
    "args": {
      "Python parent id": 802, "Python id": 804, "finished": true, "Ev Idx": 7634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825269.761, "dur": 0.268,
    "args": {
      "Python parent id": 802, "Python id": 805, "finished": true, "Ev Idx": 7635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_18", "pid": 537, "tid": 537,
    "ts": 7429564825272.679, "dur": 123.060,
    "args": {
      "Python parent id": 802, "Python module id": 18, "Python id": 806, "finished": true, "Ev Idx": 7636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825273.358, "dur": 121.983,
    "args": {
      "Python parent id": 806, "Python id": 807, "finished": true, "Ev Idx": 7637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825273.506, "dur": 0.291,
    "args": {
      "Python parent id": 807, "Python id": 808, "finished": true, "Ev Idx": 7638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564825274.570, "dur": 120.528,
    "args": {
      "Python parent id": 807, "Python id": 809, "finished": true, "Ev Idx": 7639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825274.833, "dur": 0.209,
    "args": {
      "Python parent id": 809, "Python id": 810, "finished": true, "Ev Idx": 7640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825275.297, "dur": 0.151,
    "args": {
      "Python parent id": 809, "Python id": 811, "finished": true, "Ev Idx": 7641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564825275.550, "dur": 119.286,
    "args": {
      "Python parent id": 809, "Python id": 812, "finished": true, "Ev Idx": 7642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_2", "pid": 537, "tid": 537,
    "ts": 7429564825398.367, "dur": 18.606,
    "args": {
      "Python parent id": 802, "Python module id": 2, "Python id": 813, "finished": true, "Ev Idx": 7643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825399.017, "dur": 17.691,
    "args": {
      "Python parent id": 813, "Python id": 814, "finished": true, "Ev Idx": 7644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825399.247, "dur": 0.495,
    "args": {
      "Python parent id": 814, "Python id": 815, "finished": true, "Ev Idx": 7645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564825400.860, "dur": 15.752,
    "args": {
      "Python parent id": 814, "Python id": 816, "finished": true, "Ev Idx": 7646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564825401.710, "dur": 14.769,
    "args": {
      "Python parent id": 816, "Python id": 817, "finished": true, "Ev Idx": 7647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564825401.890, "dur": 0.308,
    "args": {
      "Python parent id": 817, "Python id": 818, "finished": true, "Ev Idx": 7648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564825403.157, "dur": 13.230,
    "args": {
      "Python parent id": 817, "Python id": 819, "finished": true, "Ev Idx": 7649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825417.779, "dur": 0.508,
    "args": {
      "Python parent id": 802, "Python id": 820, "finished": true, "Ev Idx": 7650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_19", "pid": 537, "tid": 537,
    "ts": 7429564825421.105, "dur": 88.251,
    "args": {
      "Python parent id": 802, "Python module id": 19, "Python id": 821, "finished": true, "Ev Idx": 7651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825421.626, "dur": 87.482,
    "args": {
      "Python parent id": 821, "Python id": 822, "finished": true, "Ev Idx": 7652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825421.836, "dur": 0.219,
    "args": {
      "Python parent id": 822, "Python id": 823, "finished": true, "Ev Idx": 7653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564825422.909, "dur": 86.050,
    "args": {
      "Python parent id": 822, "Python id": 824, "finished": true, "Ev Idx": 7654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825423.191, "dur": 0.130,
    "args": {
      "Python parent id": 824, "Python id": 825, "finished": true, "Ev Idx": 7655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825423.682, "dur": 0.136,
    "args": {
      "Python parent id": 824, "Python id": 826, "finished": true, "Ev Idx": 7656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564825423.944, "dur": 84.838,
    "args": {
      "Python parent id": 824, "Python id": 827, "finished": true, "Ev Idx": 7657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_20", "pid": 537, "tid": 537,
    "ts": 7429564825523.436, "dur": 45.963,
    "args": {
      "Python parent id": 802, "Python module id": 20, "Python id": 828, "finished": true, "Ev Idx": 7658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825524.193, "dur": 44.868,
    "args": {
      "Python parent id": 828, "Python id": 829, "finished": true, "Ev Idx": 7659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825524.426, "dur": 0.424,
    "args": {
      "Python parent id": 829, "Python id": 830, "finished": true, "Ev Idx": 7660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564825525.784, "dur": 43.142,
    "args": {
      "Python parent id": 829, "Python id": 831, "finished": true, "Ev Idx": 7661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825526.073, "dur": 0.200,
    "args": {
      "Python parent id": 831, "Python id": 832, "finished": true, "Ev Idx": 7662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825526.460, "dur": 0.123,
    "args": {
      "Python parent id": 831, "Python id": 833, "finished": true, "Ev Idx": 7663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564825526.699, "dur": 42.179,
    "args": {
      "Python parent id": 831, "Python id": 834, "finished": true, "Ev Idx": 7664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_3", "pid": 537, "tid": 537,
    "ts": 7429564825591.439, "dur": 1249.046,
    "args": {
      "Python parent id": 64, "Python module id": 3, "Python id": 835, "finished": true, "Ev Idx": 7665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825592.908, "dur": 1246.823,
    "args": {
      "Python parent id": 835, "Python id": 836, "finished": true, "Ev Idx": 7666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825593.140, "dur": 0.405,
    "args": {
      "Python parent id": 836, "Python id": 837, "finished": true, "Ev Idx": 7667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564825595.175, "dur": 1243.728,
    "args": {
      "Python parent id": 836, "Python id": 838, "finished": true, "Ev Idx": 7668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825595.707, "dur": 0.525,
    "args": {
      "Python parent id": 838, "Python id": 839, "finished": true, "Ev Idx": 7669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_6", "pid": 537, "tid": 537,
    "ts": 7429564825598.465, "dur": 82.910,
    "args": {
      "Python parent id": 838, "Python module id": 6, "Python id": 840, "finished": true, "Ev Idx": 7670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825599.023, "dur": 82.063,
    "args": {
      "Python parent id": 840, "Python id": 841, "finished": true, "Ev Idx": 7671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825599.166, "dur": 0.134,
    "args": {
      "Python parent id": 841, "Python id": 842, "finished": true, "Ev Idx": 7672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564825600.125, "dur": 80.051,
    "args": {
      "Python parent id": 841, "Python id": 843, "finished": true, "Ev Idx": 7673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564825600.803, "dur": 2.410,
    "args": {
      "Python parent id": 843, "Python id": 844, "finished": true, "Ev Idx": 7674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564825603.527, "dur": 13.021,
    "args": {
      "Python parent id": 843, "Python id": 845, "finished": true, "Ev Idx": 7675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564825616.922, "dur": 15.459,
    "args": {
      "Python parent id": 843, "Python id": 846, "finished": true, "Ev Idx": 7676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564825644.934, "dur": 11.942,
    "args": {
      "Python parent id": 843, "Python id": 847, "finished": true, "Ev Idx": 7677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825669.881, "dur": 0.194,
    "args": {
      "Python parent id": 843, "Python id": 848, "finished": true, "Ev Idx": 7678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564825670.471, "dur": 1.399,
    "args": {
      "Python parent id": 843, "Python id": 849, "finished": true, "Ev Idx": 7679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825682.009, "dur": 0.314,
    "args": {
      "Python parent id": 838, "Python id": 850, "finished": true, "Ev Idx": 7680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_3", "pid": 537, "tid": 537,
    "ts": 7429564825689.927, "dur": 751.411,
    "args": {
      "Python parent id": 838, "Python module id": 3, "Python id": 851, "finished": true, "Ev Idx": 7681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825691.315, "dur": 749.448,
    "args": {
      "Python parent id": 851, "Python id": 852, "finished": true, "Ev Idx": 7682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825691.473, "dur": 0.293,
    "args": {
      "Python parent id": 852, "Python id": 853, "finished": true, "Ev Idx": 7683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564825693.324, "dur": 746.585,
    "args": {
      "Python parent id": 852, "Python id": 854, "finished": true, "Ev Idx": 7684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825696.587, "dur": 0.370,
    "args": {
      "Python parent id": 854, "Python id": 855, "finished": true, "Ev Idx": 7685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_21", "pid": 537, "tid": 537,
    "ts": 7429564825698.974, "dur": 40.093,
    "args": {
      "Python parent id": 854, "Python module id": 21, "Python id": 856, "finished": true, "Ev Idx": 7686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825699.778, "dur": 39.034,
    "args": {
      "Python parent id": 856, "Python id": 857, "finished": true, "Ev Idx": 7687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825699.926, "dur": 0.208,
    "args": {
      "Python parent id": 857, "Python id": 858, "finished": true, "Ev Idx": 7688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564825700.793, "dur": 37.871,
    "args": {
      "Python parent id": 857, "Python id": 859, "finished": true, "Ev Idx": 7689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825701.154, "dur": 0.146,
    "args": {
      "Python parent id": 859, "Python id": 860, "finished": true, "Ev Idx": 7690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825701.542, "dur": 0.139,
    "args": {
      "Python parent id": 859, "Python id": 861, "finished": true, "Ev Idx": 7691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564825701.855, "dur": 36.708,
    "args": {
      "Python parent id": 859, "Python id": 862, "finished": true, "Ev Idx": 7692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564825739.755, "dur": 3.650,
    "args": {
      "Python parent id": 854, "Python id": 863, "finished": true, "Ev Idx": 7693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564825744.333, "dur": 4.107,
    "args": {
      "Python parent id": 854, "Python id": 864, "finished": true, "Ev Idx": 7694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825749.344, "dur": 0.363,
    "args": {
      "Python parent id": 854, "Python id": 865, "finished": true, "Ev Idx": 7695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_22", "pid": 537, "tid": 537,
    "ts": 7429564825751.894, "dur": 43.934,
    "args": {
      "Python parent id": 854, "Python module id": 22, "Python id": 866, "finished": true, "Ev Idx": 7696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825752.503, "dur": 43.139,
    "args": {
      "Python parent id": 866, "Python id": 867, "finished": true, "Ev Idx": 7697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825752.712, "dur": 0.227,
    "args": {
      "Python parent id": 867, "Python id": 868, "finished": true, "Ev Idx": 7698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564825753.581, "dur": 41.958,
    "args": {
      "Python parent id": 867, "Python id": 869, "finished": true, "Ev Idx": 7699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825753.834, "dur": 0.148,
    "args": {
      "Python parent id": 869, "Python id": 870, "finished": true, "Ev Idx": 7700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825754.147, "dur": 0.124,
    "args": {
      "Python parent id": 869, "Python id": 871, "finished": true, "Ev Idx": 7701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564825754.376, "dur": 41.085,
    "args": {
      "Python parent id": 869, "Python id": 872, "finished": true, "Ev Idx": 7702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564825796.272, "dur": 2.627,
    "args": {
      "Python parent id": 854, "Python id": 873, "finished": true, "Ev Idx": 7703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564825799.466, "dur": 2.812,
    "args": {
      "Python parent id": 854, "Python id": 874, "finished": true, "Ev Idx": 7704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825803.100, "dur": 0.426,
    "args": {
      "Python parent id": 854, "Python id": 875, "finished": true, "Ev Idx": 7705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_23", "pid": 537, "tid": 537,
    "ts": 7429564825806.054, "dur": 37.442,
    "args": {
      "Python parent id": 854, "Python module id": 23, "Python id": 876, "finished": true, "Ev Idx": 7706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564825806.645, "dur": 36.681,
    "args": {
      "Python parent id": 876, "Python id": 877, "finished": true, "Ev Idx": 7707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564825806.832, "dur": 0.273,
    "args": {
      "Python parent id": 877, "Python id": 878, "finished": true, "Ev Idx": 7708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564825807.811, "dur": 35.295,
    "args": {
      "Python parent id": 877, "Python id": 879, "finished": true, "Ev Idx": 7709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825808.060, "dur": 0.216,
    "args": {
      "Python parent id": 879, "Python id": 880, "finished": true, "Ev Idx": 7710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564825808.503, "dur": 0.111,
    "args": {
      "Python parent id": 879, "Python id": 881, "finished": true, "Ev Idx": 7711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564825808.717, "dur": 34.309,
    "args": {
      "Python parent id": 879, "Python id": 882, "finished": true, "Ev Idx": 7712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564825844.043, "dur": 2.511,
    "args": {
      "Python parent id": 854, "Python id": 883, "finished": true, "Ev Idx": 7713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564825847.014, "dur": 2.543,
    "args": {
      "Python parent id": 854, "Python id": 884, "finished": true, "Ev Idx": 7714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564825850.773, "dur": 198.335,
    "args": {
      "Python parent id": 854, "Python id": 885, "finished": true, "Ev Idx": 7715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564825851.070, "dur": 3.527,
    "args": {
      "Python parent id": 885, "Python id": 886, "finished": true, "Ev Idx": 7716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564825854.906, "dur": 1.842,
    "args": {
      "Python parent id": 885, "Python id": 887, "finished": true, "Ev Idx": 7717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564825868.048, "dur": 95.000,
    "args": {
      "Python parent id": 885, "Python id": 888, "finished": true, "Ev Idx": 7718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564825898.420, "dur": 63.574,
    "args": {
      "Python parent id": 888, "Python id": 889, "finished": true, "Ev Idx": 7719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564825996.399, "dur": 32.895,
    "args": {
      "Python parent id": 885, "Python id": 890, "finished": true, "Ev Idx": 7720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564826015.662, "dur": 12.824,
    "args": {
      "Python parent id": 890, "Python id": 891, "finished": true, "Ev Idx": 7721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564826054.190, "dur": 3.857,
    "args": {
      "Python parent id": 854, "Python id": 892, "finished": true, "Ev Idx": 7722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564826055.015, "dur": 0.307,
    "args": {
      "Python parent id": 892, "Python id": 893, "finished": true, "Ev Idx": 7723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564826055.731, "dur": 0.067,
    "args": {
      "Python parent id": 892, "Python id": 894, "finished": true, "Ev Idx": 7724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564826057.046, "dur": 0.278,
    "args": {
      "Python parent id": 892, "Python id": 895, "finished": true, "Ev Idx": 7725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564826057.689, "dur": 0.109,
    "args": {
      "Python parent id": 892, "Python id": 896, "finished": true, "Ev Idx": 7726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564826059.449, "dur": 6.863,
    "args": {
      "Python parent id": 854, "Python id": 897, "finished": true, "Ev Idx": 7727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564826062.771, "dur": 3.391,
    "args": {
      "Python parent id": 897, "Python id": 898, "finished": true, "Ev Idx": 7728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564826063.061, "dur": 1.494,
    "args": {
      "Python parent id": 898, "Python id": 899, "finished": true, "Ev Idx": 7729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564826063.670, "dur": 0.699,
    "args": {
      "Python parent id": 899, "Python id": 900, "finished": true, "Ev Idx": 7730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564826064.822, "dur": 0.522,
    "args": {
      "Python parent id": 898, "Python id": 901, "finished": true, "Ev Idx": 7731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564826065.656, "dur": 0.434,
    "args": {
      "Python parent id": 898, "Python id": 902, "finished": true, "Ev Idx": 7732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564826069.178, "dur": 309.521,
    "args": {
      "Python parent id": 854, "Python id": 903, "finished": true, "Ev Idx": 7733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564826087.399, "dur": 57.096,
    "args": {
      "Python parent id": 903, "Python id": 904, "finished": true, "Ev Idx": 7734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564826109.547, "dur": 6.419,
    "args": {
      "Python parent id": 904, "Python id": 905, "finished": true, "Ev Idx": 7735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564826117.014, "dur": 27.331,
    "args": {
      "Python parent id": 904, "Python id": 906, "finished": true, "Ev Idx": 7736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564826145.613, "dur": 34.520,
    "args": {
      "Python parent id": 903, "Python id": 907, "finished": true, "Ev Idx": 7737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564826162.412, "dur": 2.828,
    "args": {
      "Python parent id": 907, "Python id": 908, "finished": true, "Ev Idx": 7738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564826165.827, "dur": 14.243,
    "args": {
      "Python parent id": 907, "Python id": 909, "finished": true, "Ev Idx": 7739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564826181.169, "dur": 3.684,
    "args": {
      "Python parent id": 903, "Python id": 910, "finished": true, "Ev Idx": 7740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564826185.156, "dur": 56.192,
    "args": {
      "Python parent id": 903, "Python id": 911, "finished": true, "Ev Idx": 7741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564826279.582, "dur": 17.938,
    "args": {
      "Python parent id": 903, "Python id": 912, "finished": true, "Ev Idx": 7742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564826279.786, "dur": 0.302,
    "args": {
      "Python parent id": 912, "Python id": 913, "finished": true, "Ev Idx": 7743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564826280.899, "dur": 16.459,
    "args": {
      "Python parent id": 912, "Python id": 914, "finished": true, "Ev Idx": 7744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564826298.588, "dur": 1.998,
    "args": {
      "Python parent id": 903, "Python id": 915, "finished": true, "Ev Idx": 7745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564826302.259, "dur": 11.678,
    "args": {
      "Python parent id": 903, "Python id": 916, "finished": true, "Ev Idx": 7746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564826302.382, "dur": 0.113,
    "args": {
      "Python parent id": 916, "Python id": 917, "finished": true, "Ev Idx": 7747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826308.651, "dur": 1.648,
    "args": {
      "Python parent id": 916, "Python id": 918, "finished": true, "Ev Idx": 7748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564826309.703, "dur": 0.530,
    "args": {
      "Python parent id": 918, "Python id": 919, "finished": true, "Ev Idx": 7749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564826310.697, "dur": 3.153,
    "args": {
      "Python parent id": 916, "Python id": 920, "finished": true, "Ev Idx": 7750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564826314.699, "dur": 44.312,
    "args": {
      "Python parent id": 903, "Python id": 921, "finished": true, "Ev Idx": 7751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564826359.649, "dur": 3.141,
    "args": {
      "Python parent id": 903, "Python id": 922, "finished": true, "Ev Idx": 7752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564826363.287, "dur": 14.398,
    "args": {
      "Python parent id": 903, "Python id": 923, "finished": true, "Ev Idx": 7753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564826381.755, "dur": 3.620,
    "args": {
      "Python parent id": 854, "Python id": 924, "finished": true, "Ev Idx": 7754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564826386.103, "dur": 0.276,
    "args": {
      "Python parent id": 854, "Python id": 925, "finished": true, "Ev Idx": 7755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826387.487, "dur": 1.113,
    "args": {
      "Python parent id": 854, "Python id": 926, "finished": true, "Ev Idx": 7756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_24", "pid": 537, "tid": 537,
    "ts": 7429564826394.280, "dur": 45.017,
    "args": {
      "Python parent id": 854, "Python module id": 24, "Python id": 927, "finished": true, "Ev Idx": 7757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826395.450, "dur": 43.278,
    "args": {
      "Python parent id": 927, "Python id": 928, "finished": true, "Ev Idx": 7758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826396.037, "dur": 1.432,
    "args": {
      "Python parent id": 928, "Python id": 929, "finished": true, "Ev Idx": 7759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564826398.741, "dur": 39.714,
    "args": {
      "Python parent id": 928, "Python id": 930, "finished": true, "Ev Idx": 7760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826399.253, "dur": 0.226,
    "args": {
      "Python parent id": 930, "Python id": 931, "finished": true, "Ev Idx": 7761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826399.773, "dur": 0.150,
    "args": {
      "Python parent id": 930, "Python id": 932, "finished": true, "Ev Idx": 7762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564826400.074, "dur": 38.278,
    "args": {
      "Python parent id": 930, "Python id": 933, "finished": true, "Ev Idx": 7763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826454.211, "dur": 0.550,
    "args": {
      "Python parent id": 838, "Python id": 934, "finished": true, "Ev Idx": 7764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_7", "pid": 537, "tid": 537,
    "ts": 7429564826457.745, "dur": 85.829,
    "args": {
      "Python parent id": 838, "Python module id": 7, "Python id": 935, "finished": true, "Ev Idx": 7765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826458.361, "dur": 84.879,
    "args": {
      "Python parent id": 935, "Python id": 936, "finished": true, "Ev Idx": 7766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826458.621, "dur": 0.442,
    "args": {
      "Python parent id": 936, "Python id": 937, "finished": true, "Ev Idx": 7767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564826460.033, "dur": 82.355,
    "args": {
      "Python parent id": 936, "Python id": 938, "finished": true, "Ev Idx": 7768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564826460.563, "dur": 1.705,
    "args": {
      "Python parent id": 938, "Python id": 939, "finished": true, "Ev Idx": 7769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564826462.598, "dur": 14.716,
    "args": {
      "Python parent id": 938, "Python id": 940, "finished": true, "Ev Idx": 7770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564826477.686, "dur": 15.719,
    "args": {
      "Python parent id": 938, "Python id": 941, "finished": true, "Ev Idx": 7771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564826507.685, "dur": 11.306,
    "args": {
      "Python parent id": 938, "Python id": 942, "finished": true, "Ev Idx": 7772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826529.589, "dur": 0.270,
    "args": {
      "Python parent id": 938, "Python id": 943, "finished": true, "Ev Idx": 7773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564826530.240, "dur": 1.215,
    "args": {
      "Python parent id": 938, "Python id": 944, "finished": true, "Ev Idx": 7774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826544.273, "dur": 0.347,
    "args": {
      "Python parent id": 838, "Python id": 945, "finished": true, "Ev Idx": 7775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_3", "pid": 537, "tid": 537,
    "ts": 7429564826546.446, "dur": 281.514,
    "args": {
      "Python parent id": 838, "Python module id": 3, "Python id": 946, "finished": true, "Ev Idx": 7776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826546.857, "dur": 280.942,
    "args": {
      "Python parent id": 946, "Python id": 947, "finished": true, "Ev Idx": 7777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826547.087, "dur": 0.262,
    "args": {
      "Python parent id": 947, "Python id": 948, "finished": true, "Ev Idx": 7778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564826548.299, "dur": 279.371,
    "args": {
      "Python parent id": 947, "Python id": 949, "finished": true, "Ev Idx": 7779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826548.576, "dur": 0.361,
    "args": {
      "Python parent id": 949, "Python id": 950, "finished": true, "Ev Idx": 7780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826549.258, "dur": 0.292,
    "args": {
      "Python parent id": 949, "Python id": 951, "finished": true, "Ev Idx": 7781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826549.781, "dur": 0.269,
    "args": {
      "Python parent id": 949, "Python id": 952, "finished": true, "Ev Idx": 7782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_25", "pid": 537, "tid": 537,
    "ts": 7429564826552.223, "dur": 51.678,
    "args": {
      "Python parent id": 949, "Python module id": 25, "Python id": 953, "finished": true, "Ev Idx": 7783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826552.824, "dur": 50.830,
    "args": {
      "Python parent id": 953, "Python id": 954, "finished": true, "Ev Idx": 7784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826552.959, "dur": 0.167,
    "args": {
      "Python parent id": 954, "Python id": 955, "finished": true, "Ev Idx": 7785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564826553.891, "dur": 49.632,
    "args": {
      "Python parent id": 954, "Python id": 956, "finished": true, "Ev Idx": 7786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826554.228, "dur": 0.149,
    "args": {
      "Python parent id": 956, "Python id": 957, "finished": true, "Ev Idx": 7787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826554.597, "dur": 0.151,
    "args": {
      "Python parent id": 956, "Python id": 958, "finished": true, "Ev Idx": 7788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564826554.864, "dur": 48.546,
    "args": {
      "Python parent id": 956, "Python id": 959, "finished": true, "Ev Idx": 7789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_3", "pid": 537, "tid": 537,
    "ts": 7429564826606.267, "dur": 74.273,
    "args": {
      "Python parent id": 949, "Python module id": 3, "Python id": 960, "finished": true, "Ev Idx": 7790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826606.898, "dur": 73.370,
    "args": {
      "Python parent id": 960, "Python id": 961, "finished": true, "Ev Idx": 7791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826607.099, "dur": 0.321,
    "args": {
      "Python parent id": 961, "Python id": 962, "finished": true, "Ev Idx": 7792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564826608.568, "dur": 71.535,
    "args": {
      "Python parent id": 961, "Python id": 963, "finished": true, "Ev Idx": 7793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564826609.540, "dur": 70.381,
    "args": {
      "Python parent id": 963, "Python id": 964, "finished": true, "Ev Idx": 7794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564826609.653, "dur": 0.206,
    "args": {
      "Python parent id": 964, "Python id": 965, "finished": true, "Ev Idx": 7795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564826610.808, "dur": 68.984,
    "args": {
      "Python parent id": 964, "Python id": 966, "finished": true, "Ev Idx": 7796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826681.257, "dur": 0.475,
    "args": {
      "Python parent id": 949, "Python id": 967, "finished": true, "Ev Idx": 7797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_26", "pid": 537, "tid": 537,
    "ts": 7429564826684.033, "dur": 51.554,
    "args": {
      "Python parent id": 949, "Python module id": 26, "Python id": 968, "finished": true, "Ev Idx": 7798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826684.619, "dur": 50.774,
    "args": {
      "Python parent id": 968, "Python id": 969, "finished": true, "Ev Idx": 7799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826684.824, "dur": 0.340,
    "args": {
      "Python parent id": 969, "Python id": 970, "finished": true, "Ev Idx": 7800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564826686.078, "dur": 49.100,
    "args": {
      "Python parent id": 969, "Python id": 971, "finished": true, "Ev Idx": 7801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826686.368, "dur": 0.282,
    "args": {
      "Python parent id": 971, "Python id": 972, "finished": true, "Ev Idx": 7802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826686.836, "dur": 0.123,
    "args": {
      "Python parent id": 971, "Python id": 973, "finished": true, "Ev Idx": 7803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564826687.063, "dur": 48.013,
    "args": {
      "Python parent id": 971, "Python id": 974, "finished": true, "Ev Idx": 7804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_27", "pid": 537, "tid": 537,
    "ts": 7429564826778.656, "dur": 48.684,
    "args": {
      "Python parent id": 949, "Python module id": 27, "Python id": 975, "finished": true, "Ev Idx": 7805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826779.548, "dur": 47.435,
    "args": {
      "Python parent id": 975, "Python id": 976, "finished": true, "Ev Idx": 7806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826779.833, "dur": 0.421,
    "args": {
      "Python parent id": 976, "Python id": 977, "finished": true, "Ev Idx": 7807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564826781.052, "dur": 45.792,
    "args": {
      "Python parent id": 976, "Python id": 978, "finished": true, "Ev Idx": 7808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826781.332, "dur": 0.240,
    "args": {
      "Python parent id": 978, "Python id": 979, "finished": true, "Ev Idx": 7809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826781.821, "dur": 0.133,
    "args": {
      "Python parent id": 978, "Python id": 980, "finished": true, "Ev Idx": 7810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564826782.064, "dur": 44.704,
    "args": {
      "Python parent id": 978, "Python id": 981, "finished": true, "Ev Idx": 7811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_4", "pid": 537, "tid": 537,
    "ts": 7429564826846.913, "dur": 1193.858,
    "args": {
      "Python parent id": 64, "Python module id": 4, "Python id": 982, "finished": true, "Ev Idx": 7812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826848.319, "dur": 1191.649,
    "args": {
      "Python parent id": 982, "Python id": 983, "finished": true, "Ev Idx": 7813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826848.565, "dur": 0.416,
    "args": {
      "Python parent id": 983, "Python id": 984, "finished": true, "Ev Idx": 7814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564826850.835, "dur": 1188.384,
    "args": {
      "Python parent id": 983, "Python id": 985, "finished": true, "Ev Idx": 7815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826851.333, "dur": 0.445,
    "args": {
      "Python parent id": 985, "Python id": 986, "finished": true, "Ev Idx": 7816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_8", "pid": 537, "tid": 537,
    "ts": 7429564826853.688, "dur": 78.624,
    "args": {
      "Python parent id": 985, "Python module id": 8, "Python id": 987, "finished": true, "Ev Idx": 7817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826854.230, "dur": 77.849,
    "args": {
      "Python parent id": 987, "Python id": 988, "finished": true, "Ev Idx": 7818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826854.466, "dur": 0.149,
    "args": {
      "Python parent id": 988, "Python id": 989, "finished": true, "Ev Idx": 7819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564826855.594, "dur": 75.675,
    "args": {
      "Python parent id": 988, "Python id": 990, "finished": true, "Ev Idx": 7820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564826856.135, "dur": 2.263,
    "args": {
      "Python parent id": 990, "Python id": 991, "finished": true, "Ev Idx": 7821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564826858.691, "dur": 11.730,
    "args": {
      "Python parent id": 990, "Python id": 992, "finished": true, "Ev Idx": 7822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564826870.839, "dur": 15.281,
    "args": {
      "Python parent id": 990, "Python id": 993, "finished": true, "Ev Idx": 7823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564826898.917, "dur": 10.580,
    "args": {
      "Python parent id": 990, "Python id": 994, "finished": true, "Ev Idx": 7824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826920.493, "dur": 0.168,
    "args": {
      "Python parent id": 990, "Python id": 995, "finished": true, "Ev Idx": 7825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564826921.034, "dur": 1.363,
    "args": {
      "Python parent id": 990, "Python id": 996, "finished": true, "Ev Idx": 7826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826933.049, "dur": 0.302,
    "args": {
      "Python parent id": 985, "Python id": 997, "finished": true, "Ev Idx": 7827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_4", "pid": 537, "tid": 537,
    "ts": 7429564826936.243, "dur": 684.162,
    "args": {
      "Python parent id": 985, "Python module id": 4, "Python id": 998, "finished": true, "Ev Idx": 7828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826937.523, "dur": 682.511,
    "args": {
      "Python parent id": 998, "Python id": 999, "finished": true, "Ev Idx": 7829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826937.796, "dur": 0.301,
    "args": {
      "Python parent id": 999, "Python id": 1000, "finished": true, "Ev Idx": 7830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564826939.702, "dur": 679.448,
    "args": {
      "Python parent id": 999, "Python id": 1001, "finished": true, "Ev Idx": 7831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826942.695, "dur": 0.314,
    "args": {
      "Python parent id": 1001, "Python id": 1002, "finished": true, "Ev Idx": 7832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_28", "pid": 537, "tid": 537,
    "ts": 7429564826945.418, "dur": 42.961,
    "args": {
      "Python parent id": 1001, "Python module id": 28, "Python id": 1003, "finished": true, "Ev Idx": 7833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564826945.933, "dur": 42.183,
    "args": {
      "Python parent id": 1003, "Python id": 1004, "finished": true, "Ev Idx": 7834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564826946.188, "dur": 0.157,
    "args": {
      "Python parent id": 1004, "Python id": 1005, "finished": true, "Ev Idx": 7835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564826947.198, "dur": 40.803,
    "args": {
      "Python parent id": 1004, "Python id": 1006, "finished": true, "Ev Idx": 7836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826947.486, "dur": 0.135,
    "args": {
      "Python parent id": 1006, "Python id": 1007, "finished": true, "Ev Idx": 7837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826947.836, "dur": 0.383,
    "args": {
      "Python parent id": 1006, "Python id": 1008, "finished": true, "Ev Idx": 7838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564826948.343, "dur": 39.560,
    "args": {
      "Python parent id": 1006, "Python id": 1009, "finished": true, "Ev Idx": 7839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564826989.280, "dur": 3.616,
    "args": {
      "Python parent id": 1001, "Python id": 1010, "finished": true, "Ev Idx": 7840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564826993.365, "dur": 4.122,
    "args": {
      "Python parent id": 1001, "Python id": 1011, "finished": true, "Ev Idx": 7841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564826998.280, "dur": 0.369,
    "args": {
      "Python parent id": 1001, "Python id": 1012, "finished": true, "Ev Idx": 7842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_29", "pid": 537, "tid": 537,
    "ts": 7429564827000.922, "dur": 42.815,
    "args": {
      "Python parent id": 1001, "Python module id": 29, "Python id": 1013, "finished": true, "Ev Idx": 7843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564827001.446, "dur": 42.073,
    "args": {
      "Python parent id": 1013, "Python id": 1014, "finished": true, "Ev Idx": 7844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564827001.672, "dur": 0.269,
    "args": {
      "Python parent id": 1014, "Python id": 1015, "finished": true, "Ev Idx": 7845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564827002.718, "dur": 40.697,
    "args": {
      "Python parent id": 1014, "Python id": 1016, "finished": true, "Ev Idx": 7846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827002.966, "dur": 0.161,
    "args": {
      "Python parent id": 1016, "Python id": 1017, "finished": true, "Ev Idx": 7847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827003.335, "dur": 0.157,
    "args": {
      "Python parent id": 1016, "Python id": 1018, "finished": true, "Ev Idx": 7848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564827003.614, "dur": 39.751,
    "args": {
      "Python parent id": 1016, "Python id": 1019, "finished": true, "Ev Idx": 7849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564827044.174, "dur": 2.540,
    "args": {
      "Python parent id": 1001, "Python id": 1020, "finished": true, "Ev Idx": 7850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564827047.200, "dur": 2.976,
    "args": {
      "Python parent id": 1001, "Python id": 1021, "finished": true, "Ev Idx": 7851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827050.929, "dur": 0.481,
    "args": {
      "Python parent id": 1001, "Python id": 1022, "finished": true, "Ev Idx": 7852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_30", "pid": 537, "tid": 537,
    "ts": 7429564827053.726, "dur": 42.047,
    "args": {
      "Python parent id": 1001, "Python module id": 30, "Python id": 1023, "finished": true, "Ev Idx": 7853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564827054.326, "dur": 41.199,
    "args": {
      "Python parent id": 1023, "Python id": 1024, "finished": true, "Ev Idx": 7854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564827054.553, "dur": 0.252,
    "args": {
      "Python parent id": 1024, "Python id": 1025, "finished": true, "Ev Idx": 7855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564827055.479, "dur": 39.944,
    "args": {
      "Python parent id": 1024, "Python id": 1026, "finished": true, "Ev Idx": 7856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827055.684, "dur": 0.128,
    "args": {
      "Python parent id": 1026, "Python id": 1027, "finished": true, "Ev Idx": 7857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827056.008, "dur": 0.114,
    "args": {
      "Python parent id": 1026, "Python id": 1028, "finished": true, "Ev Idx": 7858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564827056.225, "dur": 39.149,
    "args": {
      "Python parent id": 1026, "Python id": 1029, "finished": true, "Ev Idx": 7859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564827096.270, "dur": 2.425,
    "args": {
      "Python parent id": 1001, "Python id": 1030, "finished": true, "Ev Idx": 7860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564827099.194, "dur": 2.806,
    "args": {
      "Python parent id": 1001, "Python id": 1031, "finished": true, "Ev Idx": 7861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564827103.211, "dur": 147.116,
    "args": {
      "Python parent id": 1001, "Python id": 1032, "finished": true, "Ev Idx": 7862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564827103.467, "dur": 3.796,
    "args": {
      "Python parent id": 1032, "Python id": 1033, "finished": true, "Ev Idx": 7863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564827107.552, "dur": 1.772,
    "args": {
      "Python parent id": 1032, "Python id": 1034, "finished": true, "Ev Idx": 7864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564827121.056, "dur": 50.875,
    "args": {
      "Python parent id": 1032, "Python id": 1035, "finished": true, "Ev Idx": 7865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564827148.097, "dur": 22.648,
    "args": {
      "Python parent id": 1035, "Python id": 1036, "finished": true, "Ev Idx": 7866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564827202.382, "dur": 28.853,
    "args": {
      "Python parent id": 1032, "Python id": 1037, "finished": true, "Ev Idx": 7867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564827219.396, "dur": 11.126,
    "args": {
      "Python parent id": 1037, "Python id": 1038, "finished": true, "Ev Idx": 7868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564827254.657, "dur": 4.011,
    "args": {
      "Python parent id": 1001, "Python id": 1039, "finished": true, "Ev Idx": 7869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564827255.321, "dur": 0.363,
    "args": {
      "Python parent id": 1039, "Python id": 1040, "finished": true, "Ev Idx": 7870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564827256.010, "dur": 0.070,
    "args": {
      "Python parent id": 1039, "Python id": 1041, "finished": true, "Ev Idx": 7871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564827257.393, "dur": 0.377,
    "args": {
      "Python parent id": 1039, "Python id": 1042, "finished": true, "Ev Idx": 7872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564827258.267, "dur": 0.132,
    "args": {
      "Python parent id": 1039, "Python id": 1043, "finished": true, "Ev Idx": 7873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564827260.038, "dur": 6.712,
    "args": {
      "Python parent id": 1001, "Python id": 1044, "finished": true, "Ev Idx": 7874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564827263.218, "dur": 3.411,
    "args": {
      "Python parent id": 1044, "Python id": 1045, "finished": true, "Ev Idx": 7875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564827263.474, "dur": 1.552,
    "args": {
      "Python parent id": 1045, "Python id": 1046, "finished": true, "Ev Idx": 7876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564827264.129, "dur": 0.745,
    "args": {
      "Python parent id": 1046, "Python id": 1047, "finished": true, "Ev Idx": 7877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564827265.341, "dur": 0.454,
    "args": {
      "Python parent id": 1045, "Python id": 1048, "finished": true, "Ev Idx": 7878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564827266.099, "dur": 0.452,
    "args": {
      "Python parent id": 1045, "Python id": 1049, "finished": true, "Ev Idx": 7879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564827269.485, "dur": 290.432,
    "args": {
      "Python parent id": 1001, "Python id": 1050, "finished": true, "Ev Idx": 7880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564827270.036, "dur": 78.131,
    "args": {
      "Python parent id": 1050, "Python id": 1051, "finished": true, "Ev Idx": 7881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564827321.316, "dur": 5.040,
    "args": {
      "Python parent id": 1051, "Python id": 1052, "finished": true, "Ev Idx": 7882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564827327.186, "dur": 20.849,
    "args": {
      "Python parent id": 1051, "Python id": 1053, "finished": true, "Ev Idx": 7883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564827349.018, "dur": 30.199,
    "args": {
      "Python parent id": 1050, "Python id": 1054, "finished": true, "Ev Idx": 7884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564827362.758, "dur": 2.612,
    "args": {
      "Python parent id": 1054, "Python id": 1055, "finished": true, "Ev Idx": 7885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564827365.886, "dur": 13.239,
    "args": {
      "Python parent id": 1054, "Python id": 1056, "finished": true, "Ev Idx": 7886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564827380.228, "dur": 3.045,
    "args": {
      "Python parent id": 1050, "Python id": 1057, "finished": true, "Ev Idx": 7887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564827383.578, "dur": 48.947,
    "args": {
      "Python parent id": 1050, "Python id": 1058, "finished": true, "Ev Idx": 7888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564827466.075, "dur": 16.371,
    "args": {
      "Python parent id": 1050, "Python id": 1059, "finished": true, "Ev Idx": 7889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564827466.269, "dur": 0.294,
    "args": {
      "Python parent id": 1059, "Python id": 1060, "finished": true, "Ev Idx": 7890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564827467.247, "dur": 15.071,
    "args": {
      "Python parent id": 1059, "Python id": 1061, "finished": true, "Ev Idx": 7891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564827483.379, "dur": 1.760,
    "args": {
      "Python parent id": 1050, "Python id": 1062, "finished": true, "Ev Idx": 7892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564827486.643, "dur": 9.901,
    "args": {
      "Python parent id": 1050, "Python id": 1063, "finished": true, "Ev Idx": 7893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564827486.826, "dur": 0.075,
    "args": {
      "Python parent id": 1063, "Python id": 1064, "finished": true, "Ev Idx": 7894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827491.980, "dur": 1.366,
    "args": {
      "Python parent id": 1063, "Python id": 1065, "finished": true, "Ev Idx": 7895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564827492.856, "dur": 0.411,
    "args": {
      "Python parent id": 1065, "Python id": 1066, "finished": true, "Ev Idx": 7896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564827493.768, "dur": 2.699,
    "args": {
      "Python parent id": 1063, "Python id": 1067, "finished": true, "Ev Idx": 7897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564827497.200, "dur": 44.095,
    "args": {
      "Python parent id": 1050, "Python id": 1068, "finished": true, "Ev Idx": 7898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564827541.919, "dur": 3.050,
    "args": {
      "Python parent id": 1050, "Python id": 1069, "finished": true, "Ev Idx": 7899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564827545.434, "dur": 13.414,
    "args": {
      "Python parent id": 1050, "Python id": 1070, "finished": true, "Ev Idx": 7900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564827562.907, "dur": 3.403,
    "args": {
      "Python parent id": 1001, "Python id": 1071, "finished": true, "Ev Idx": 7901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564827567.029, "dur": 0.254,
    "args": {
      "Python parent id": 1001, "Python id": 1072, "finished": true, "Ev Idx": 7902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827568.290, "dur": 1.019,
    "args": {
      "Python parent id": 1001, "Python id": 1073, "finished": true, "Ev Idx": 7903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_31", "pid": 537, "tid": 537,
    "ts": 7429564827574.183, "dur": 44.303,
    "args": {
      "Python parent id": 1001, "Python module id": 31, "Python id": 1074, "finished": true, "Ev Idx": 7904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564827575.294, "dur": 42.680,
    "args": {
      "Python parent id": 1074, "Python id": 1075, "finished": true, "Ev Idx": 7905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564827575.738, "dur": 1.199,
    "args": {
      "Python parent id": 1075, "Python id": 1076, "finished": true, "Ev Idx": 7906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564827578.340, "dur": 39.327,
    "args": {
      "Python parent id": 1075, "Python id": 1077, "finished": true, "Ev Idx": 7907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827578.794, "dur": 0.312,
    "args": {
      "Python parent id": 1077, "Python id": 1078, "finished": true, "Ev Idx": 7908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827579.379, "dur": 0.150,
    "args": {
      "Python parent id": 1077, "Python id": 1079, "finished": true, "Ev Idx": 7909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564827579.724, "dur": 37.818,
    "args": {
      "Python parent id": 1077, "Python id": 1080, "finished": true, "Ev Idx": 7910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827632.297, "dur": 0.553,
    "args": {
      "Python parent id": 985, "Python id": 1081, "finished": true, "Ev Idx": 7911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_9", "pid": 537, "tid": 537,
    "ts": 7429564827635.699, "dur": 150.142,
    "args": {
      "Python parent id": 985, "Python module id": 9, "Python id": 1082, "finished": true, "Ev Idx": 7912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564827636.524, "dur": 148.852,
    "args": {
      "Python parent id": 1082, "Python id": 1083, "finished": true, "Ev Idx": 7913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564827636.760, "dur": 0.440,
    "args": {
      "Python parent id": 1083, "Python id": 1084, "finished": true, "Ev Idx": 7914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564827638.370, "dur": 145.570,
    "args": {
      "Python parent id": 1083, "Python id": 1085, "finished": true, "Ev Idx": 7915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564827638.859, "dur": 1.548,
    "args": {
      "Python parent id": 1085, "Python id": 1086, "finished": true, "Ev Idx": 7916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564827640.709, "dur": 13.728,
    "args": {
      "Python parent id": 1085, "Python id": 1087, "finished": true, "Ev Idx": 7917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564827654.891, "dur": 15.709,
    "args": {
      "Python parent id": 1085, "Python id": 1088, "finished": true, "Ev Idx": 7918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564827683.731, "dur": 10.799,
    "args": {
      "Python parent id": 1085, "Python id": 1089, "finished": true, "Ev Idx": 7919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827706.956, "dur": 0.361,
    "args": {
      "Python parent id": 1085, "Python id": 1090, "finished": true, "Ev Idx": 7920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564827707.724, "dur": 1.145,
    "args": {
      "Python parent id": 1085, "Python id": 1091, "finished": true, "Ev Idx": 7921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827786.846, "dur": 0.525,
    "args": {
      "Python parent id": 985, "Python id": 1092, "finished": true, "Ev Idx": 7922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_4", "pid": 537, "tid": 537,
    "ts": 7429564827793.608, "dur": 233.957,
    "args": {
      "Python parent id": 985, "Python module id": 4, "Python id": 1093, "finished": true, "Ev Idx": 7923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564827794.246, "dur": 233.066,
    "args": {
      "Python parent id": 1093, "Python id": 1094, "finished": true, "Ev Idx": 7924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564827794.552, "dur": 0.707,
    "args": {
      "Python parent id": 1094, "Python id": 1095, "finished": true, "Ev Idx": 7925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564827796.357, "dur": 230.801,
    "args": {
      "Python parent id": 1094, "Python id": 1096, "finished": true, "Ev Idx": 7926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827796.646, "dur": 0.453,
    "args": {
      "Python parent id": 1096, "Python id": 1097, "finished": true, "Ev Idx": 7927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827797.347, "dur": 0.321,
    "args": {
      "Python parent id": 1096, "Python id": 1098, "finished": true, "Ev Idx": 7928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827797.961, "dur": 0.262,
    "args": {
      "Python parent id": 1096, "Python id": 1099, "finished": true, "Ev Idx": 7929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_32", "pid": 537, "tid": 537,
    "ts": 7429564827800.954, "dur": 60.379,
    "args": {
      "Python parent id": 1096, "Python module id": 32, "Python id": 1100, "finished": true, "Ev Idx": 7930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564827801.600, "dur": 59.446,
    "args": {
      "Python parent id": 1100, "Python id": 1101, "finished": true, "Ev Idx": 7931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564827801.747, "dur": 0.158,
    "args": {
      "Python parent id": 1101, "Python id": 1102, "finished": true, "Ev Idx": 7932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564827802.690, "dur": 58.181,
    "args": {
      "Python parent id": 1101, "Python id": 1103, "finished": true, "Ev Idx": 7933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827803.073, "dur": 0.176,
    "args": {
      "Python parent id": 1103, "Python id": 1104, "finished": true, "Ev Idx": 7934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827803.539, "dur": 0.147,
    "args": {
      "Python parent id": 1103, "Python id": 1105, "finished": true, "Ev Idx": 7935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564827803.859, "dur": 56.913,
    "args": {
      "Python parent id": 1103, "Python id": 1106, "finished": true, "Ev Idx": 7936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_4", "pid": 537, "tid": 537,
    "ts": 7429564827863.401, "dur": 17.903,
    "args": {
      "Python parent id": 1096, "Python module id": 4, "Python id": 1107, "finished": true, "Ev Idx": 7937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564827864.117, "dur": 16.897,
    "args": {
      "Python parent id": 1107, "Python id": 1108, "finished": true, "Ev Idx": 7938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564827864.318, "dur": 0.322,
    "args": {
      "Python parent id": 1108, "Python id": 1109, "finished": true, "Ev Idx": 7939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564827865.912, "dur": 14.993,
    "args": {
      "Python parent id": 1108, "Python id": 1110, "finished": true, "Ev Idx": 7940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564827866.909, "dur": 13.843,
    "args": {
      "Python parent id": 1110, "Python id": 1111, "finished": true, "Ev Idx": 7941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564827867.059, "dur": 0.257,
    "args": {
      "Python parent id": 1111, "Python id": 1112, "finished": true, "Ev Idx": 7942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564827868.310, "dur": 12.339,
    "args": {
      "Python parent id": 1111, "Python id": 1113, "finished": true, "Ev Idx": 7943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827882.039, "dur": 0.343,
    "args": {
      "Python parent id": 1096, "Python id": 1114, "finished": true, "Ev Idx": 7944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_33", "pid": 537, "tid": 537,
    "ts": 7429564827884.665, "dur": 85.397,
    "args": {
      "Python parent id": 1096, "Python module id": 33, "Python id": 1115, "finished": true, "Ev Idx": 7945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564827885.441, "dur": 84.396,
    "args": {
      "Python parent id": 1115, "Python id": 1116, "finished": true, "Ev Idx": 7946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564827885.638, "dur": 0.251,
    "args": {
      "Python parent id": 1116, "Python id": 1117, "finished": true, "Ev Idx": 7947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564827886.639, "dur": 83.015,
    "args": {
      "Python parent id": 1116, "Python id": 1118, "finished": true, "Ev Idx": 7948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827886.847, "dur": 0.129,
    "args": {
      "Python parent id": 1118, "Python id": 1119, "finished": true, "Ev Idx": 7949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827887.206, "dur": 0.155,
    "args": {
      "Python parent id": 1118, "Python id": 1120, "finished": true, "Ev Idx": 7950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564827887.498, "dur": 82.007,
    "args": {
      "Python parent id": 1118, "Python id": 1121, "finished": true, "Ev Idx": 7951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_34", "pid": 537, "tid": 537,
    "ts": 7429564827983.110, "dur": 43.727,
    "args": {
      "Python parent id": 1096, "Python module id": 34, "Python id": 1122, "finished": true, "Ev Idx": 7952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564827983.986, "dur": 42.609,
    "args": {
      "Python parent id": 1122, "Python id": 1123, "finished": true, "Ev Idx": 7953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564827984.279, "dur": 0.392,
    "args": {
      "Python parent id": 1123, "Python id": 1124, "finished": true, "Ev Idx": 7954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564827985.449, "dur": 41.035,
    "args": {
      "Python parent id": 1123, "Python id": 1125, "finished": true, "Ev Idx": 7955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827985.691, "dur": 0.180,
    "args": {
      "Python parent id": 1125, "Python id": 1126, "finished": true, "Ev Idx": 7956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564827986.090, "dur": 0.163,
    "args": {
      "Python parent id": 1125, "Python id": 1127, "finished": true, "Ev Idx": 7957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564827986.378, "dur": 40.053,
    "args": {
      "Python parent id": 1125, "Python id": 1128, "finished": true, "Ev Idx": 7958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_5", "pid": 537, "tid": 537,
    "ts": 7429564828047.922, "dur": 1197.195,
    "args": {
      "Python parent id": 64, "Python module id": 5, "Python id": 1129, "finished": true, "Ev Idx": 7959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564828049.469, "dur": 1194.773,
    "args": {
      "Python parent id": 1129, "Python id": 1130, "finished": true, "Ev Idx": 7960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564828049.708, "dur": 0.389,
    "args": {
      "Python parent id": 1130, "Python id": 1131, "finished": true, "Ev Idx": 7961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564828051.921, "dur": 1191.434,
    "args": {
      "Python parent id": 1130, "Python id": 1132, "finished": true, "Ev Idx": 7962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828052.577, "dur": 0.371,
    "args": {
      "Python parent id": 1132, "Python id": 1133, "finished": true, "Ev Idx": 7963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_10", "pid": 537, "tid": 537,
    "ts": 7429564828055.112, "dur": 80.791,
    "args": {
      "Python parent id": 1132, "Python module id": 10, "Python id": 1134, "finished": true, "Ev Idx": 7964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564828055.622, "dur": 80.011,
    "args": {
      "Python parent id": 1134, "Python id": 1135, "finished": true, "Ev Idx": 7965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564828055.769, "dur": 0.171,
    "args": {
      "Python parent id": 1135, "Python id": 1136, "finished": true, "Ev Idx": 7966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564828056.787, "dur": 77.787,
    "args": {
      "Python parent id": 1135, "Python id": 1137, "finished": true, "Ev Idx": 7967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564828057.334, "dur": 2.014,
    "args": {
      "Python parent id": 1137, "Python id": 1138, "finished": true, "Ev Idx": 7968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564828059.624, "dur": 12.945,
    "args": {
      "Python parent id": 1137, "Python id": 1139, "finished": true, "Ev Idx": 7969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564828072.901, "dur": 16.246,
    "args": {
      "Python parent id": 1137, "Python id": 1140, "finished": true, "Ev Idx": 7970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564828101.964, "dur": 9.284,
    "args": {
      "Python parent id": 1137, "Python id": 1141, "finished": true, "Ev Idx": 7971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828122.808, "dur": 0.238,
    "args": {
      "Python parent id": 1137, "Python id": 1142, "finished": true, "Ev Idx": 7972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564828123.521, "dur": 1.248,
    "args": {
      "Python parent id": 1137, "Python id": 1143, "finished": true, "Ev Idx": 7973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828136.513, "dur": 0.313,
    "args": {
      "Python parent id": 1132, "Python id": 1144, "finished": true, "Ev Idx": 7974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_5", "pid": 537, "tid": 537,
    "ts": 7429564828140.203, "dur": 694.353,
    "args": {
      "Python parent id": 1132, "Python module id": 5, "Python id": 1145, "finished": true, "Ev Idx": 7975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564828141.528, "dur": 692.610,
    "args": {
      "Python parent id": 1145, "Python id": 1146, "finished": true, "Ev Idx": 7976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564828141.746, "dur": 0.268,
    "args": {
      "Python parent id": 1146, "Python id": 1147, "finished": true, "Ev Idx": 7977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564828143.669, "dur": 689.778,
    "args": {
      "Python parent id": 1146, "Python id": 1148, "finished": true, "Ev Idx": 7978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828146.954, "dur": 0.464,
    "args": {
      "Python parent id": 1148, "Python id": 1149, "finished": true, "Ev Idx": 7979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_35", "pid": 537, "tid": 537,
    "ts": 7429564828150.059, "dur": 84.942,
    "args": {
      "Python parent id": 1148, "Python module id": 35, "Python id": 1150, "finished": true, "Ev Idx": 7980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564828150.539, "dur": 84.112,
    "args": {
      "Python parent id": 1150, "Python id": 1151, "finished": true, "Ev Idx": 7981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564828150.682, "dur": 0.170,
    "args": {
      "Python parent id": 1151, "Python id": 1152, "finished": true, "Ev Idx": 7982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564828151.615, "dur": 82.867,
    "args": {
      "Python parent id": 1151, "Python id": 1153, "finished": true, "Ev Idx": 7983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828151.941, "dur": 0.135,
    "args": {
      "Python parent id": 1153, "Python id": 1154, "finished": true, "Ev Idx": 7984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828152.336, "dur": 0.141,
    "args": {
      "Python parent id": 1153, "Python id": 1155, "finished": true, "Ev Idx": 7985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564828152.621, "dur": 81.779,
    "args": {
      "Python parent id": 1153, "Python id": 1156, "finished": true, "Ev Idx": 7986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564828235.763, "dur": 3.928,
    "args": {
      "Python parent id": 1148, "Python id": 1157, "finished": true, "Ev Idx": 7987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564828240.239, "dur": 4.289,
    "args": {
      "Python parent id": 1148, "Python id": 1158, "finished": true, "Ev Idx": 7988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828245.434, "dur": 0.435,
    "args": {
      "Python parent id": 1148, "Python id": 1159, "finished": true, "Ev Idx": 7989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_36", "pid": 537, "tid": 537,
    "ts": 7429564828248.715, "dur": 49.373,
    "args": {
      "Python parent id": 1148, "Python module id": 36, "Python id": 1160, "finished": true, "Ev Idx": 7990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564828249.340, "dur": 48.464,
    "args": {
      "Python parent id": 1160, "Python id": 1161, "finished": true, "Ev Idx": 7991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564828249.608, "dur": 0.391,
    "args": {
      "Python parent id": 1161, "Python id": 1162, "finished": true, "Ev Idx": 7992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564828250.775, "dur": 46.918,
    "args": {
      "Python parent id": 1161, "Python id": 1163, "finished": true, "Ev Idx": 7993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828251.094, "dur": 0.168,
    "args": {
      "Python parent id": 1163, "Python id": 1164, "finished": true, "Ev Idx": 7994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828251.535, "dur": 0.166,
    "args": {
      "Python parent id": 1163, "Python id": 1165, "finished": true, "Ev Idx": 7995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564828251.843, "dur": 45.795,
    "args": {
      "Python parent id": 1163, "Python id": 1166, "finished": true, "Ev Idx": 7996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564828298.594, "dur": 2.717,
    "args": {
      "Python parent id": 1148, "Python id": 1167, "finished": true, "Ev Idx": 7997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564828301.993, "dur": 3.078,
    "args": {
      "Python parent id": 1148, "Python id": 1168, "finished": true, "Ev Idx": 7998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828305.817, "dur": 0.524,
    "args": {
      "Python parent id": 1148, "Python id": 1169, "finished": true, "Ev Idx": 7999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_37", "pid": 537, "tid": 537,
    "ts": 7429564828308.960, "dur": 44.416,
    "args": {
      "Python parent id": 1148, "Python module id": 37, "Python id": 1170, "finished": true, "Ev Idx": 8000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564828309.521, "dur": 43.555,
    "args": {
      "Python parent id": 1170, "Python id": 1171, "finished": true, "Ev Idx": 8001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564828309.723, "dur": 0.272,
    "args": {
      "Python parent id": 1171, "Python id": 1172, "finished": true, "Ev Idx": 8002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564828310.673, "dur": 42.304,
    "args": {
      "Python parent id": 1171, "Python id": 1173, "finished": true, "Ev Idx": 8003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828310.946, "dur": 0.159,
    "args": {
      "Python parent id": 1173, "Python id": 1174, "finished": true, "Ev Idx": 8004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828311.314, "dur": 0.150,
    "args": {
      "Python parent id": 1173, "Python id": 1175, "finished": true, "Ev Idx": 8005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564828311.601, "dur": 41.329,
    "args": {
      "Python parent id": 1173, "Python id": 1176, "finished": true, "Ev Idx": 8006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564828353.860, "dur": 2.359,
    "args": {
      "Python parent id": 1148, "Python id": 1177, "finished": true, "Ev Idx": 8007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564828356.750, "dur": 2.927,
    "args": {
      "Python parent id": 1148, "Python id": 1178, "finished": true, "Ev Idx": 8008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564828360.806, "dur": 147.301,
    "args": {
      "Python parent id": 1148, "Python id": 1179, "finished": true, "Ev Idx": 8009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564828361.075, "dur": 3.599,
    "args": {
      "Python parent id": 1179, "Python id": 1180, "finished": true, "Ev Idx": 8010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564828364.964, "dur": 1.744,
    "args": {
      "Python parent id": 1179, "Python id": 1181, "finished": true, "Ev Idx": 8011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564828380.794, "dur": 49.002,
    "args": {
      "Python parent id": 1179, "Python id": 1182, "finished": true, "Ev Idx": 8012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564828407.269, "dur": 21.443,
    "args": {
      "Python parent id": 1182, "Python id": 1183, "finished": true, "Ev Idx": 8013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564828459.403, "dur": 30.335,
    "args": {
      "Python parent id": 1179, "Python id": 1184, "finished": true, "Ev Idx": 8014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564828477.658, "dur": 11.220,
    "args": {
      "Python parent id": 1184, "Python id": 1185, "finished": true, "Ev Idx": 8015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564828511.976, "dur": 3.988,
    "args": {
      "Python parent id": 1148, "Python id": 1186, "finished": true, "Ev Idx": 8016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564828512.798, "dur": 0.331,
    "args": {
      "Python parent id": 1186, "Python id": 1187, "finished": true, "Ev Idx": 8017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564828513.502, "dur": 0.036,
    "args": {
      "Python parent id": 1186, "Python id": 1188, "finished": true, "Ev Idx": 8018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564828514.898, "dur": 0.231,
    "args": {
      "Python parent id": 1186, "Python id": 1189, "finished": true, "Ev Idx": 8019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564828515.554, "dur": 0.123,
    "args": {
      "Python parent id": 1186, "Python id": 1190, "finished": true, "Ev Idx": 8020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564828517.268, "dur": 6.679,
    "args": {
      "Python parent id": 1148, "Python id": 1191, "finished": true, "Ev Idx": 8021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564828520.435, "dur": 3.342,
    "args": {
      "Python parent id": 1191, "Python id": 1192, "finished": true, "Ev Idx": 8022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564828520.704, "dur": 1.484,
    "args": {
      "Python parent id": 1192, "Python id": 1193, "finished": true, "Ev Idx": 8023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564828521.345, "dur": 0.716,
    "args": {
      "Python parent id": 1193, "Python id": 1194, "finished": true, "Ev Idx": 8024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564828522.462, "dur": 0.479,
    "args": {
      "Python parent id": 1192, "Python id": 1195, "finished": true, "Ev Idx": 8025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564828523.241, "dur": 0.462,
    "args": {
      "Python parent id": 1192, "Python id": 1196, "finished": true, "Ev Idx": 8026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564828526.606, "dur": 251.183,
    "args": {
      "Python parent id": 1148, "Python id": 1197, "finished": true, "Ev Idx": 8027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564828527.174, "dur": 45.921,
    "args": {
      "Python parent id": 1197, "Python id": 1198, "finished": true, "Ev Idx": 8028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564828545.363, "dur": 4.988,
    "args": {
      "Python parent id": 1198, "Python id": 1199, "finished": true, "Ev Idx": 8029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564828551.198, "dur": 21.777,
    "args": {
      "Python parent id": 1198, "Python id": 1200, "finished": true, "Ev Idx": 8030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564828574.043, "dur": 29.787,
    "args": {
      "Python parent id": 1197, "Python id": 1201, "finished": true, "Ev Idx": 8031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564828588.133, "dur": 3.151,
    "args": {
      "Python parent id": 1201, "Python id": 1202, "finished": true, "Ev Idx": 8032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564828591.726, "dur": 12.018,
    "args": {
      "Python parent id": 1201, "Python id": 1203, "finished": true, "Ev Idx": 8033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564828604.717, "dur": 2.917,
    "args": {
      "Python parent id": 1197, "Python id": 1204, "finished": true, "Ev Idx": 8034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564828607.905, "dur": 46.705,
    "args": {
      "Python parent id": 1197, "Python id": 1205, "finished": true, "Ev Idx": 8035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564828686.172, "dur": 16.507,
    "args": {
      "Python parent id": 1197, "Python id": 1206, "finished": true, "Ev Idx": 8036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564828686.414, "dur": 0.328,
    "args": {
      "Python parent id": 1206, "Python id": 1207, "finished": true, "Ev Idx": 8037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564828687.416, "dur": 15.142,
    "args": {
      "Python parent id": 1206, "Python id": 1208, "finished": true, "Ev Idx": 8038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564828703.548, "dur": 1.847,
    "args": {
      "Python parent id": 1197, "Python id": 1209, "finished": true, "Ev Idx": 8039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564828706.959, "dur": 10.114,
    "args": {
      "Python parent id": 1197, "Python id": 1210, "finished": true, "Ev Idx": 8040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564828707.222, "dur": 0.113,
    "args": {
      "Python parent id": 1210, "Python id": 1211, "finished": true, "Ev Idx": 8041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828712.471, "dur": 1.500,
    "args": {
      "Python parent id": 1210, "Python id": 1212, "finished": true, "Ev Idx": 8042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564828713.474, "dur": 0.399,
    "args": {
      "Python parent id": 1212, "Python id": 1213, "finished": true, "Ev Idx": 8043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564828714.379, "dur": 2.617,
    "args": {
      "Python parent id": 1210, "Python id": 1214, "finished": true, "Ev Idx": 8044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564828717.709, "dur": 41.308,
    "args": {
      "Python parent id": 1197, "Python id": 1215, "finished": true, "Ev Idx": 8045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564828759.610, "dur": 2.902,
    "args": {
      "Python parent id": 1197, "Python id": 1216, "finished": true, "Ev Idx": 8046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564828763.095, "dur": 13.699,
    "args": {
      "Python parent id": 1197, "Python id": 1217, "finished": true, "Ev Idx": 8047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564828780.541, "dur": 3.379,
    "args": {
      "Python parent id": 1148, "Python id": 1218, "finished": true, "Ev Idx": 8048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564828784.593, "dur": 0.311,
    "args": {
      "Python parent id": 1148, "Python id": 1219, "finished": true, "Ev Idx": 8049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828785.861, "dur": 0.822,
    "args": {
      "Python parent id": 1148, "Python id": 1220, "finished": true, "Ev Idx": 8050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_38", "pid": 537, "tid": 537,
    "ts": 7429564828790.344, "dur": 42.602,
    "args": {
      "Python parent id": 1148, "Python module id": 38, "Python id": 1221, "finished": true, "Ev Idx": 8051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564828791.297, "dur": 41.219,
    "args": {
      "Python parent id": 1221, "Python id": 1222, "finished": true, "Ev Idx": 8052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564828791.669, "dur": 0.906,
    "args": {
      "Python parent id": 1222, "Python id": 1223, "finished": true, "Ev Idx": 8053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564828793.668, "dur": 38.660,
    "args": {
      "Python parent id": 1222, "Python id": 1224, "finished": true, "Ev Idx": 8054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828794.050, "dur": 0.221,
    "args": {
      "Python parent id": 1224, "Python id": 1225, "finished": true, "Ev Idx": 8055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828794.540, "dur": 0.280,
    "args": {
      "Python parent id": 1224, "Python id": 1226, "finished": true, "Ev Idx": 8056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564828794.980, "dur": 37.256,
    "args": {
      "Python parent id": 1224, "Python id": 1227, "finished": true, "Ev Idx": 8057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828846.374, "dur": 0.613,
    "args": {
      "Python parent id": 1132, "Python id": 1228, "finished": true, "Ev Idx": 8058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_11", "pid": 537, "tid": 537,
    "ts": 7429564828849.803, "dur": 83.120,
    "args": {
      "Python parent id": 1132, "Python module id": 11, "Python id": 1229, "finished": true, "Ev Idx": 8059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564828850.609, "dur": 81.913,
    "args": {
      "Python parent id": 1229, "Python id": 1230, "finished": true, "Ev Idx": 8060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564828850.991, "dur": 0.390,
    "args": {
      "Python parent id": 1230, "Python id": 1231, "finished": true, "Ev Idx": 8061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564828852.686, "dur": 79.018,
    "args": {
      "Python parent id": 1230, "Python id": 1232, "finished": true, "Ev Idx": 8062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564828853.098, "dur": 1.495,
    "args": {
      "Python parent id": 1232, "Python id": 1233, "finished": true, "Ev Idx": 8063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564828854.891, "dur": 12.388,
    "args": {
      "Python parent id": 1232, "Python id": 1234, "finished": true, "Ev Idx": 8064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564828867.666, "dur": 16.543,
    "args": {
      "Python parent id": 1232, "Python id": 1235, "finished": true, "Ev Idx": 8065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564828896.452, "dur": 10.075,
    "args": {
      "Python parent id": 1232, "Python id": 1236, "finished": true, "Ev Idx": 8066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828920.114, "dur": 0.187,
    "args": {
      "Python parent id": 1232, "Python id": 1237, "finished": true, "Ev Idx": 8067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564828920.637, "dur": 1.299,
    "args": {
      "Python parent id": 1232, "Python id": 1238, "finished": true, "Ev Idx": 8068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828933.626, "dur": 0.348,
    "args": {
      "Python parent id": 1132, "Python id": 1239, "finished": true, "Ev Idx": 8069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_5", "pid": 537, "tid": 537,
    "ts": 7429564828935.590, "dur": 296.304,
    "args": {
      "Python parent id": 1132, "Python module id": 5, "Python id": 1240, "finished": true, "Ev Idx": 8070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564828936.057, "dur": 295.632,
    "args": {
      "Python parent id": 1240, "Python id": 1241, "finished": true, "Ev Idx": 8071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564828936.278, "dur": 0.442,
    "args": {
      "Python parent id": 1241, "Python id": 1242, "finished": true, "Ev Idx": 8072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564828937.702, "dur": 293.866,
    "args": {
      "Python parent id": 1241, "Python id": 1243, "finished": true, "Ev Idx": 8073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828937.943, "dur": 0.364,
    "args": {
      "Python parent id": 1243, "Python id": 1244, "finished": true, "Ev Idx": 8074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828938.704, "dur": 0.337,
    "args": {
      "Python parent id": 1243, "Python id": 1245, "finished": true, "Ev Idx": 8075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828939.297, "dur": 0.228,
    "args": {
      "Python parent id": 1243, "Python id": 1246, "finished": true, "Ev Idx": 8076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_39", "pid": 537, "tid": 537,
    "ts": 7429564828941.587, "dur": 115.433,
    "args": {
      "Python parent id": 1243, "Python module id": 39, "Python id": 1247, "finished": true, "Ev Idx": 8077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564828942.189, "dur": 114.463,
    "args": {
      "Python parent id": 1247, "Python id": 1248, "finished": true, "Ev Idx": 8078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564828942.332, "dur": 0.179,
    "args": {
      "Python parent id": 1248, "Python id": 1249, "finished": true, "Ev Idx": 8079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564828943.125, "dur": 113.183,
    "args": {
      "Python parent id": 1248, "Python id": 1250, "finished": true, "Ev Idx": 8080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828943.490, "dur": 0.134,
    "args": {
      "Python parent id": 1250, "Python id": 1251, "finished": true, "Ev Idx": 8081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564828943.794, "dur": 0.137,
    "args": {
      "Python parent id": 1250, "Python id": 1252, "finished": true, "Ev Idx": 8082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564828944.047, "dur": 112.055,
    "args": {
      "Python parent id": 1250, "Python id": 1253, "finished": true, "Ev Idx": 8083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_5", "pid": 537, "tid": 537,
    "ts": 7429564829059.573, "dur": 18.673,
    "args": {
      "Python parent id": 1243, "Python module id": 5, "Python id": 1254, "finished": true, "Ev Idx": 8084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564829060.474, "dur": 17.554,
    "args": {
      "Python parent id": 1254, "Python id": 1255, "finished": true, "Ev Idx": 8085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564829060.790, "dur": 0.567,
    "args": {
      "Python parent id": 1255, "Python id": 1256, "finished": true, "Ev Idx": 8086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564829062.452, "dur": 15.480,
    "args": {
      "Python parent id": 1255, "Python id": 1257, "finished": true, "Ev Idx": 8087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564829063.415, "dur": 14.370,
    "args": {
      "Python parent id": 1257, "Python id": 1258, "finished": true, "Ev Idx": 8088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564829063.704, "dur": 0.221,
    "args": {
      "Python parent id": 1258, "Python id": 1259, "finished": true, "Ev Idx": 8089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564829064.830, "dur": 12.859,
    "args": {
      "Python parent id": 1258, "Python id": 1260, "finished": true, "Ev Idx": 8090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829079.077, "dur": 0.615,
    "args": {
      "Python parent id": 1243, "Python id": 1261, "finished": true, "Ev Idx": 8091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_40", "pid": 537, "tid": 537,
    "ts": 7429564829081.884, "dur": 85.589,
    "args": {
      "Python parent id": 1243, "Python module id": 40, "Python id": 1262, "finished": true, "Ev Idx": 8092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564829082.682, "dur": 84.572,
    "args": {
      "Python parent id": 1262, "Python id": 1263, "finished": true, "Ev Idx": 8093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564829082.871, "dur": 0.225,
    "args": {
      "Python parent id": 1263, "Python id": 1264, "finished": true, "Ev Idx": 8094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564829083.907, "dur": 83.158,
    "args": {
      "Python parent id": 1263, "Python id": 1265, "finished": true, "Ev Idx": 8095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829084.260, "dur": 0.143,
    "args": {
      "Python parent id": 1265, "Python id": 1266, "finished": true, "Ev Idx": 8096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829084.690, "dur": 0.107,
    "args": {
      "Python parent id": 1265, "Python id": 1267, "finished": true, "Ev Idx": 8097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564829084.925, "dur": 82.017,
    "args": {
      "Python parent id": 1265, "Python id": 1268, "finished": true, "Ev Idx": 8098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_41", "pid": 537, "tid": 537,
    "ts": 7429564829181.685, "dur": 49.593,
    "args": {
      "Python parent id": 1243, "Python module id": 41, "Python id": 1269, "finished": true, "Ev Idx": 8099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564829182.412, "dur": 48.625,
    "args": {
      "Python parent id": 1269, "Python id": 1270, "finished": true, "Ev Idx": 8100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564829182.679, "dur": 0.498,
    "args": {
      "Python parent id": 1270, "Python id": 1271, "finished": true, "Ev Idx": 8101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564829184.039, "dur": 46.820,
    "args": {
      "Python parent id": 1270, "Python id": 1272, "finished": true, "Ev Idx": 8102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829184.376, "dur": 0.225,
    "args": {
      "Python parent id": 1272, "Python id": 1273, "finished": true, "Ev Idx": 8103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829184.809, "dur": 0.287,
    "args": {
      "Python parent id": 1272, "Python id": 1274, "finished": true, "Ev Idx": 8104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564829185.207, "dur": 45.577,
    "args": {
      "Python parent id": 1272, "Python id": 1275, "finished": true, "Ev Idx": 8105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_6", "pid": 537, "tid": 537,
    "ts": 7429564829251.968, "dur": 1278.561,
    "args": {
      "Python parent id": 64, "Python module id": 6, "Python id": 1276, "finished": true, "Ev Idx": 8106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564829253.347, "dur": 1276.566,
    "args": {
      "Python parent id": 1276, "Python id": 1277, "finished": true, "Ev Idx": 8107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564829253.561, "dur": 0.340,
    "args": {
      "Python parent id": 1277, "Python id": 1278, "finished": true, "Ev Idx": 8108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564829255.523, "dur": 1273.530,
    "args": {
      "Python parent id": 1277, "Python id": 1279, "finished": true, "Ev Idx": 8109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829256.050, "dur": 0.404,
    "args": {
      "Python parent id": 1279, "Python id": 1280, "finished": true, "Ev Idx": 8110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_12", "pid": 537, "tid": 537,
    "ts": 7429564829258.574, "dur": 79.696,
    "args": {
      "Python parent id": 1279, "Python module id": 12, "Python id": 1281, "finished": true, "Ev Idx": 8111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564829259.099, "dur": 78.866,
    "args": {
      "Python parent id": 1281, "Python id": 1282, "finished": true, "Ev Idx": 8112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564829259.372, "dur": 0.169,
    "args": {
      "Python parent id": 1282, "Python id": 1283, "finished": true, "Ev Idx": 8113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564829260.345, "dur": 76.862,
    "args": {
      "Python parent id": 1282, "Python id": 1284, "finished": true, "Ev Idx": 8114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564829260.915, "dur": 2.043,
    "args": {
      "Python parent id": 1284, "Python id": 1285, "finished": true, "Ev Idx": 8115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564829263.249, "dur": 12.002,
    "args": {
      "Python parent id": 1284, "Python id": 1286, "finished": true, "Ev Idx": 8116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564829275.705, "dur": 16.625,
    "args": {
      "Python parent id": 1284, "Python id": 1287, "finished": true, "Ev Idx": 8117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564829305.311, "dur": 10.004,
    "args": {
      "Python parent id": 1284, "Python id": 1288, "finished": true, "Ev Idx": 8118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829326.493, "dur": 0.207,
    "args": {
      "Python parent id": 1284, "Python id": 1289, "finished": true, "Ev Idx": 8119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564829327.105, "dur": 1.277,
    "args": {
      "Python parent id": 1284, "Python id": 1290, "finished": true, "Ev Idx": 8120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829338.892, "dur": 0.363,
    "args": {
      "Python parent id": 1279, "Python id": 1291, "finished": true, "Ev Idx": 8121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_6", "pid": 537, "tid": 537,
    "ts": 7429564829342.779, "dur": 710.868,
    "args": {
      "Python parent id": 1279, "Python module id": 6, "Python id": 1292, "finished": true, "Ev Idx": 8122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564829344.213, "dur": 708.888,
    "args": {
      "Python parent id": 1292, "Python id": 1293, "finished": true, "Ev Idx": 8123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564829344.414, "dur": 0.357,
    "args": {
      "Python parent id": 1293, "Python id": 1294, "finished": true, "Ev Idx": 8124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564829346.451, "dur": 705.865,
    "args": {
      "Python parent id": 1293, "Python id": 1295, "finished": true, "Ev Idx": 8125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829349.642, "dur": 0.376,
    "args": {
      "Python parent id": 1295, "Python id": 1296, "finished": true, "Ev Idx": 8126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_42", "pid": 537, "tid": 537,
    "ts": 7429564829352.598, "dur": 43.501,
    "args": {
      "Python parent id": 1295, "Python module id": 42, "Python id": 1297, "finished": true, "Ev Idx": 8127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564829353.182, "dur": 42.643,
    "args": {
      "Python parent id": 1297, "Python id": 1298, "finished": true, "Ev Idx": 8128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564829353.451, "dur": 0.194,
    "args": {
      "Python parent id": 1298, "Python id": 1299, "finished": true, "Ev Idx": 8129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564829354.330, "dur": 41.371,
    "args": {
      "Python parent id": 1298, "Python id": 1300, "finished": true, "Ev Idx": 8130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829354.658, "dur": 0.169,
    "args": {
      "Python parent id": 1300, "Python id": 1301, "finished": true, "Ev Idx": 8131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829355.156, "dur": 0.274,
    "args": {
      "Python parent id": 1300, "Python id": 1302, "finished": true, "Ev Idx": 8132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564829355.546, "dur": 40.081,
    "args": {
      "Python parent id": 1300, "Python id": 1303, "finished": true, "Ev Idx": 8133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564829396.728, "dur": 3.448,
    "args": {
      "Python parent id": 1295, "Python id": 1304, "finished": true, "Ev Idx": 8134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564829400.732, "dur": 3.967,
    "args": {
      "Python parent id": 1295, "Python id": 1305, "finished": true, "Ev Idx": 8135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829405.637, "dur": 0.404,
    "args": {
      "Python parent id": 1295, "Python id": 1306, "finished": true, "Ev Idx": 8136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_43", "pid": 537, "tid": 537,
    "ts": 7429564829408.546, "dur": 44.956,
    "args": {
      "Python parent id": 1295, "Python module id": 43, "Python id": 1307, "finished": true, "Ev Idx": 8137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564829409.086, "dur": 44.133,
    "args": {
      "Python parent id": 1307, "Python id": 1308, "finished": true, "Ev Idx": 8138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564829409.305, "dur": 0.249,
    "args": {
      "Python parent id": 1308, "Python id": 1309, "finished": true, "Ev Idx": 8139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564829410.344, "dur": 42.760,
    "args": {
      "Python parent id": 1308, "Python id": 1310, "finished": true, "Ev Idx": 8140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829410.682, "dur": 0.158,
    "args": {
      "Python parent id": 1310, "Python id": 1311, "finished": true, "Ev Idx": 8141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829411.016, "dur": 0.139,
    "args": {
      "Python parent id": 1310, "Python id": 1312, "finished": true, "Ev Idx": 8142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564829411.272, "dur": 41.784,
    "args": {
      "Python parent id": 1310, "Python id": 1313, "finished": true, "Ev Idx": 8143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564829453.988, "dur": 2.643,
    "args": {
      "Python parent id": 1295, "Python id": 1314, "finished": true, "Ev Idx": 8144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564829457.103, "dur": 2.881,
    "args": {
      "Python parent id": 1295, "Python id": 1315, "finished": true, "Ev Idx": 8145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829460.790, "dur": 0.524,
    "args": {
      "Python parent id": 1295, "Python id": 1316, "finished": true, "Ev Idx": 8146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_44", "pid": 537, "tid": 537,
    "ts": 7429564829463.612, "dur": 42.377,
    "args": {
      "Python parent id": 1295, "Python module id": 44, "Python id": 1317, "finished": true, "Ev Idx": 8147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564829464.189, "dur": 41.547,
    "args": {
      "Python parent id": 1317, "Python id": 1318, "finished": true, "Ev Idx": 8148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564829464.524, "dur": 0.279,
    "args": {
      "Python parent id": 1318, "Python id": 1319, "finished": true, "Ev Idx": 8149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564829465.470, "dur": 40.110,
    "args": {
      "Python parent id": 1318, "Python id": 1320, "finished": true, "Ev Idx": 8150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829465.735, "dur": 0.129,
    "args": {
      "Python parent id": 1320, "Python id": 1321, "finished": true, "Ev Idx": 8151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829466.063, "dur": 0.306,
    "args": {
      "Python parent id": 1320, "Python id": 1322, "finished": true, "Ev Idx": 8152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564829466.479, "dur": 39.030,
    "args": {
      "Python parent id": 1320, "Python id": 1323, "finished": true, "Ev Idx": 8153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564829506.775, "dur": 2.496,
    "args": {
      "Python parent id": 1295, "Python id": 1324, "finished": true, "Ev Idx": 8154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564829509.799, "dur": 2.550,
    "args": {
      "Python parent id": 1295, "Python id": 1325, "finished": true, "Ev Idx": 8155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564829513.586, "dur": 192.057,
    "args": {
      "Python parent id": 1295, "Python id": 1326, "finished": true, "Ev Idx": 8156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564829513.863, "dur": 3.600,
    "args": {
      "Python parent id": 1326, "Python id": 1327, "finished": true, "Ev Idx": 8157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564829517.793, "dur": 2.173,
    "args": {
      "Python parent id": 1326, "Python id": 1328, "finished": true, "Ev Idx": 8158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564829531.811, "dur": 50.339,
    "args": {
      "Python parent id": 1326, "Python id": 1329, "finished": true, "Ev Idx": 8159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564829559.080, "dur": 21.926,
    "args": {
      "Python parent id": 1329, "Python id": 1330, "finished": true, "Ev Idx": 8160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564829655.836, "dur": 32.262,
    "args": {
      "Python parent id": 1326, "Python id": 1331, "finished": true, "Ev Idx": 8161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564829674.176, "dur": 12.938,
    "args": {
      "Python parent id": 1331, "Python id": 1332, "finished": true, "Ev Idx": 8162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564829710.700, "dur": 4.100,
    "args": {
      "Python parent id": 1295, "Python id": 1333, "finished": true, "Ev Idx": 8163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564829711.686, "dur": 0.268,
    "args": {
      "Python parent id": 1333, "Python id": 1334, "finished": true, "Ev Idx": 8164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564829712.301, "dur": 0.066,
    "args": {
      "Python parent id": 1333, "Python id": 1335, "finished": true, "Ev Idx": 8165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564829713.715, "dur": 0.243,
    "args": {
      "Python parent id": 1333, "Python id": 1336, "finished": true, "Ev Idx": 8166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564829714.419, "dur": 0.111,
    "args": {
      "Python parent id": 1333, "Python id": 1337, "finished": true, "Ev Idx": 8167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564829716.241, "dur": 6.616,
    "args": {
      "Python parent id": 1295, "Python id": 1338, "finished": true, "Ev Idx": 8168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564829719.318, "dur": 3.332,
    "args": {
      "Python parent id": 1338, "Python id": 1339, "finished": true, "Ev Idx": 8169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564829719.577, "dur": 1.420,
    "args": {
      "Python parent id": 1339, "Python id": 1340, "finished": true, "Ev Idx": 8170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564829720.125, "dur": 0.723,
    "args": {
      "Python parent id": 1340, "Python id": 1341, "finished": true, "Ev Idx": 8171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564829721.312, "dur": 0.503,
    "args": {
      "Python parent id": 1339, "Python id": 1342, "finished": true, "Ev Idx": 8172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564829722.125, "dur": 0.443,
    "args": {
      "Python parent id": 1339, "Python id": 1343, "finished": true, "Ev Idx": 8173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564829725.712, "dur": 270.425,
    "args": {
      "Python parent id": 1295, "Python id": 1344, "finished": true, "Ev Idx": 8174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564829726.357, "dur": 48.219,
    "args": {
      "Python parent id": 1344, "Python id": 1345, "finished": true, "Ev Idx": 8175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564829745.335, "dur": 5.406,
    "args": {
      "Python parent id": 1345, "Python id": 1346, "finished": true, "Ev Idx": 8176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564829751.688, "dur": 22.764,
    "args": {
      "Python parent id": 1345, "Python id": 1347, "finished": true, "Ev Idx": 8177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564829775.501, "dur": 31.178,
    "args": {
      "Python parent id": 1344, "Python id": 1348, "finished": true, "Ev Idx": 8178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564829789.275, "dur": 3.291,
    "args": {
      "Python parent id": 1348, "Python id": 1349, "finished": true, "Ev Idx": 8179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564829793.041, "dur": 13.573,
    "args": {
      "Python parent id": 1348, "Python id": 1350, "finished": true, "Ev Idx": 8180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564829807.694, "dur": 3.313,
    "args": {
      "Python parent id": 1344, "Python id": 1351, "finished": true, "Ev Idx": 8181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564829811.379, "dur": 51.974,
    "args": {
      "Python parent id": 1344, "Python id": 1352, "finished": true, "Ev Idx": 8182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564829898.986, "dur": 18.457,
    "args": {
      "Python parent id": 1344, "Python id": 1353, "finished": true, "Ev Idx": 8183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564829899.286, "dur": 0.262,
    "args": {
      "Python parent id": 1353, "Python id": 1354, "finished": true, "Ev Idx": 8184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564829900.323, "dur": 16.966,
    "args": {
      "Python parent id": 1353, "Python id": 1355, "finished": true, "Ev Idx": 8185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564829918.387, "dur": 2.201,
    "args": {
      "Python parent id": 1344, "Python id": 1356, "finished": true, "Ev Idx": 8186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564829922.115, "dur": 10.668,
    "args": {
      "Python parent id": 1344, "Python id": 1357, "finished": true, "Ev Idx": 8187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564829922.262, "dur": 0.137,
    "args": {
      "Python parent id": 1357, "Python id": 1358, "finished": true, "Ev Idx": 8188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564829928.307, "dur": 1.265,
    "args": {
      "Python parent id": 1357, "Python id": 1359, "finished": true, "Ev Idx": 8189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564829929.149, "dur": 0.320,
    "args": {
      "Python parent id": 1359, "Python id": 1360, "finished": true, "Ev Idx": 8190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564829929.968, "dur": 2.732,
    "args": {
      "Python parent id": 1357, "Python id": 1361, "finished": true, "Ev Idx": 8191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564829933.510, "dur": 42.956,
    "args": {
      "Python parent id": 1344, "Python id": 1362, "finished": true, "Ev Idx": 8192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564829977.083, "dur": 3.360,
    "args": {
      "Python parent id": 1344, "Python id": 1363, "finished": true, "Ev Idx": 8193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564829980.956, "dur": 14.124,
    "args": {
      "Python parent id": 1344, "Python id": 1364, "finished": true, "Ev Idx": 8194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564829999.118, "dur": 3.211,
    "args": {
      "Python parent id": 1295, "Python id": 1365, "finished": true, "Ev Idx": 8195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564830002.999, "dur": 0.273,
    "args": {
      "Python parent id": 1295, "Python id": 1366, "finished": true, "Ev Idx": 8196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830004.362, "dur": 0.996,
    "args": {
      "Python parent id": 1295, "Python id": 1367, "finished": true, "Ev Idx": 8197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_45", "pid": 537, "tid": 537,
    "ts": 7429564830009.819, "dur": 41.907,
    "args": {
      "Python parent id": 1295, "Python module id": 45, "Python id": 1368, "finished": true, "Ev Idx": 8198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830010.838, "dur": 40.460,
    "args": {
      "Python parent id": 1368, "Python id": 1369, "finished": true, "Ev Idx": 8199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830011.173, "dur": 1.371,
    "args": {
      "Python parent id": 1369, "Python id": 1370, "finished": true, "Ev Idx": 8200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564830013.931, "dur": 37.149,
    "args": {
      "Python parent id": 1369, "Python id": 1371, "finished": true, "Ev Idx": 8201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830014.327, "dur": 0.212,
    "args": {
      "Python parent id": 1371, "Python id": 1372, "finished": true, "Ev Idx": 8202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830014.801, "dur": 0.165,
    "args": {
      "Python parent id": 1371, "Python id": 1373, "finished": true, "Ev Idx": 8203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564830015.119, "dur": 35.870,
    "args": {
      "Python parent id": 1371, "Python id": 1374, "finished": true, "Ev Idx": 8204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830123.347, "dur": 0.937,
    "args": {
      "Python parent id": 1279, "Python id": 1375, "finished": true, "Ev Idx": 8205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_13", "pid": 537, "tid": 537,
    "ts": 7429564830127.795, "dur": 89.327,
    "args": {
      "Python parent id": 1279, "Python module id": 13, "Python id": 1376, "finished": true, "Ev Idx": 8206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830128.631, "dur": 88.145,
    "args": {
      "Python parent id": 1376, "Python id": 1377, "finished": true, "Ev Idx": 8207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830128.886, "dur": 0.593,
    "args": {
      "Python parent id": 1377, "Python id": 1378, "finished": true, "Ev Idx": 8208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564830130.549, "dur": 85.399,
    "args": {
      "Python parent id": 1377, "Python id": 1379, "finished": true, "Ev Idx": 8209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564830131.124, "dur": 2.189,
    "args": {
      "Python parent id": 1379, "Python id": 1380, "finished": true, "Ev Idx": 8210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564830133.627, "dur": 15.224,
    "args": {
      "Python parent id": 1379, "Python id": 1381, "finished": true, "Ev Idx": 8211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564830149.262, "dur": 17.938,
    "args": {
      "Python parent id": 1379, "Python id": 1382, "finished": true, "Ev Idx": 8212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564830180.405, "dur": 11.966,
    "args": {
      "Python parent id": 1379, "Python id": 1383, "finished": true, "Ev Idx": 8213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830203.339, "dur": 0.295,
    "args": {
      "Python parent id": 1379, "Python id": 1384, "finished": true, "Ev Idx": 8214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564830204.006, "dur": 1.174,
    "args": {
      "Python parent id": 1379, "Python id": 1385, "finished": true, "Ev Idx": 8215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830217.913, "dur": 0.409,
    "args": {
      "Python parent id": 1279, "Python id": 1386, "finished": true, "Ev Idx": 8216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_6", "pid": 537, "tid": 537,
    "ts": 7429564830220.786, "dur": 296.983,
    "args": {
      "Python parent id": 1279, "Python module id": 6, "Python id": 1387, "finished": true, "Ev Idx": 8217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830221.379, "dur": 296.179,
    "args": {
      "Python parent id": 1387, "Python id": 1388, "finished": true, "Ev Idx": 8218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830221.607, "dur": 0.412,
    "args": {
      "Python parent id": 1388, "Python id": 1389, "finished": true, "Ev Idx": 8219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564830223.044, "dur": 294.305,
    "args": {
      "Python parent id": 1388, "Python id": 1390, "finished": true, "Ev Idx": 8220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830223.413, "dur": 0.413,
    "args": {
      "Python parent id": 1390, "Python id": 1391, "finished": true, "Ev Idx": 8221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830224.166, "dur": 0.282,
    "args": {
      "Python parent id": 1390, "Python id": 1392, "finished": true, "Ev Idx": 8222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830224.828, "dur": 0.217,
    "args": {
      "Python parent id": 1390, "Python id": 1393, "finished": true, "Ev Idx": 8223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_46", "pid": 537, "tid": 537,
    "ts": 7429564830227.174, "dur": 58.766,
    "args": {
      "Python parent id": 1390, "Python module id": 46, "Python id": 1394, "finished": true, "Ev Idx": 8224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830227.984, "dur": 57.620,
    "args": {
      "Python parent id": 1394, "Python id": 1395, "finished": true, "Ev Idx": 8225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830228.128, "dur": 0.218,
    "args": {
      "Python parent id": 1395, "Python id": 1396, "finished": true, "Ev Idx": 8226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564830229.187, "dur": 56.215,
    "args": {
      "Python parent id": 1395, "Python id": 1397, "finished": true, "Ev Idx": 8227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830229.588, "dur": 0.128,
    "args": {
      "Python parent id": 1397, "Python id": 1398, "finished": true, "Ev Idx": 8228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830229.958, "dur": 0.167,
    "args": {
      "Python parent id": 1397, "Python id": 1399, "finished": true, "Ev Idx": 8229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564830230.227, "dur": 55.040,
    "args": {
      "Python parent id": 1397, "Python id": 1400, "finished": true, "Ev Idx": 8230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_6", "pid": 537, "tid": 537,
    "ts": 7429564830287.914, "dur": 78.010,
    "args": {
      "Python parent id": 1390, "Python module id": 6, "Python id": 1401, "finished": true, "Ev Idx": 8231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830288.463, "dur": 77.092,
    "args": {
      "Python parent id": 1401, "Python id": 1402, "finished": true, "Ev Idx": 8232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830288.645, "dur": 0.350,
    "args": {
      "Python parent id": 1402, "Python id": 1403, "finished": true, "Ev Idx": 8233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564830290.126, "dur": 75.236,
    "args": {
      "Python parent id": 1402, "Python id": 1404, "finished": true, "Ev Idx": 8234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564830291.062, "dur": 74.116,
    "args": {
      "Python parent id": 1404, "Python id": 1405, "finished": true, "Ev Idx": 8235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564830291.195, "dur": 0.230,
    "args": {
      "Python parent id": 1405, "Python id": 1406, "finished": true, "Ev Idx": 8236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564830292.346, "dur": 72.691,
    "args": {
      "Python parent id": 1405, "Python id": 1407, "finished": true, "Ev Idx": 8237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830366.550, "dur": 0.419,
    "args": {
      "Python parent id": 1390, "Python id": 1408, "finished": true, "Ev Idx": 8238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_47", "pid": 537, "tid": 537,
    "ts": 7429564830369.500, "dur": 52.956,
    "args": {
      "Python parent id": 1390, "Python module id": 47, "Python id": 1409, "finished": true, "Ev Idx": 8239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830369.998, "dur": 52.125,
    "args": {
      "Python parent id": 1409, "Python id": 1410, "finished": true, "Ev Idx": 8240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830370.192, "dur": 0.331,
    "args": {
      "Python parent id": 1410, "Python id": 1411, "finished": true, "Ev Idx": 8241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564830371.318, "dur": 50.607,
    "args": {
      "Python parent id": 1410, "Python id": 1412, "finished": true, "Ev Idx": 8242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830371.621, "dur": 0.149,
    "args": {
      "Python parent id": 1412, "Python id": 1413, "finished": true, "Ev Idx": 8243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830372.012, "dur": 0.324,
    "args": {
      "Python parent id": 1412, "Python id": 1414, "finished": true, "Ev Idx": 8244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564830372.457, "dur": 49.351,
    "args": {
      "Python parent id": 1412, "Python id": 1415, "finished": true, "Ev Idx": 8245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_48", "pid": 537, "tid": 537,
    "ts": 7429564830467.357, "dur": 49.646,
    "args": {
      "Python parent id": 1390, "Python module id": 48, "Python id": 1416, "finished": true, "Ev Idx": 8246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830467.909, "dur": 48.864,
    "args": {
      "Python parent id": 1416, "Python id": 1417, "finished": true, "Ev Idx": 8247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830468.109, "dur": 0.455,
    "args": {
      "Python parent id": 1417, "Python id": 1418, "finished": true, "Ev Idx": 8248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564830469.289, "dur": 47.274,
    "args": {
      "Python parent id": 1417, "Python id": 1419, "finished": true, "Ev Idx": 8249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830469.581, "dur": 0.180,
    "args": {
      "Python parent id": 1419, "Python id": 1420, "finished": true, "Ev Idx": 8250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830470.021, "dur": 0.133,
    "args": {
      "Python parent id": 1419, "Python id": 1421, "finished": true, "Ev Idx": 8251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564830470.270, "dur": 46.253,
    "args": {
      "Python parent id": 1419, "Python id": 1422, "finished": true, "Ev Idx": 8252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_7", "pid": 537, "tid": 537,
    "ts": 7429564830537.448, "dur": 1173.034,
    "args": {
      "Python parent id": 64, "Python module id": 7, "Python id": 1423, "finished": true, "Ev Idx": 8253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830538.752, "dur": 1171.042,
    "args": {
      "Python parent id": 1423, "Python id": 1424, "finished": true, "Ev Idx": 8254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830538.989, "dur": 0.397,
    "args": {
      "Python parent id": 1424, "Python id": 1425, "finished": true, "Ev Idx": 8255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564830541.070, "dur": 1168.010,
    "args": {
      "Python parent id": 1424, "Python id": 1426, "finished": true, "Ev Idx": 8256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830541.634, "dur": 0.656,
    "args": {
      "Python parent id": 1426, "Python id": 1427, "finished": true, "Ev Idx": 8257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_14", "pid": 537, "tid": 537,
    "ts": 7429564830544.670, "dur": 78.931,
    "args": {
      "Python parent id": 1426, "Python module id": 14, "Python id": 1428, "finished": true, "Ev Idx": 8258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830545.152, "dur": 78.244,
    "args": {
      "Python parent id": 1428, "Python id": 1429, "finished": true, "Ev Idx": 8259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830545.304, "dur": 0.305,
    "args": {
      "Python parent id": 1429, "Python id": 1430, "finished": true, "Ev Idx": 8260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564830546.371, "dur": 76.281,
    "args": {
      "Python parent id": 1429, "Python id": 1431, "finished": true, "Ev Idx": 8261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564830546.996, "dur": 1.786,
    "args": {
      "Python parent id": 1431, "Python id": 1432, "finished": true, "Ev Idx": 8262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564830549.057, "dur": 11.238,
    "args": {
      "Python parent id": 1431, "Python id": 1433, "finished": true, "Ev Idx": 8263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564830560.782, "dur": 14.268,
    "args": {
      "Python parent id": 1431, "Python id": 1434, "finished": true, "Ev Idx": 8264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564830588.798, "dur": 10.844,
    "args": {
      "Python parent id": 1431, "Python id": 1435, "finished": true, "Ev Idx": 8265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830610.554, "dur": 0.300,
    "args": {
      "Python parent id": 1431, "Python id": 1436, "finished": true, "Ev Idx": 8266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564830611.464, "dur": 1.219,
    "args": {
      "Python parent id": 1431, "Python id": 1437, "finished": true, "Ev Idx": 8267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830624.312, "dur": 0.416,
    "args": {
      "Python parent id": 1426, "Python id": 1438, "finished": true, "Ev Idx": 8268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_7", "pid": 537, "tid": 537,
    "ts": 7429564830627.672, "dur": 662.101,
    "args": {
      "Python parent id": 1426, "Python module id": 7, "Python id": 1439, "finished": true, "Ev Idx": 8269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830628.883, "dur": 660.410,
    "args": {
      "Python parent id": 1439, "Python id": 1440, "finished": true, "Ev Idx": 8270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830629.076, "dur": 0.293,
    "args": {
      "Python parent id": 1440, "Python id": 1441, "finished": true, "Ev Idx": 8271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564830630.888, "dur": 657.699,
    "args": {
      "Python parent id": 1440, "Python id": 1442, "finished": true, "Ev Idx": 8272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830634.361, "dur": 0.374,
    "args": {
      "Python parent id": 1442, "Python id": 1443, "finished": true, "Ev Idx": 8273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_49", "pid": 537, "tid": 537,
    "ts": 7429564830637.169, "dur": 44.067,
    "args": {
      "Python parent id": 1442, "Python module id": 49, "Python id": 1444, "finished": true, "Ev Idx": 8274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830637.801, "dur": 43.194,
    "args": {
      "Python parent id": 1444, "Python id": 1445, "finished": true, "Ev Idx": 8275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830637.967, "dur": 0.182,
    "args": {
      "Python parent id": 1445, "Python id": 1446, "finished": true, "Ev Idx": 8276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564830638.905, "dur": 41.936,
    "args": {
      "Python parent id": 1445, "Python id": 1447, "finished": true, "Ev Idx": 8277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830639.238, "dur": 0.151,
    "args": {
      "Python parent id": 1447, "Python id": 1448, "finished": true, "Ev Idx": 8278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830639.613, "dur": 0.140,
    "args": {
      "Python parent id": 1447, "Python id": 1449, "finished": true, "Ev Idx": 8279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564830639.864, "dur": 40.896,
    "args": {
      "Python parent id": 1447, "Python id": 1450, "finished": true, "Ev Idx": 8280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564830681.943, "dur": 3.508,
    "args": {
      "Python parent id": 1442, "Python id": 1451, "finished": true, "Ev Idx": 8281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564830686.071, "dur": 3.735,
    "args": {
      "Python parent id": 1442, "Python id": 1452, "finished": true, "Ev Idx": 8282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830690.745, "dur": 0.347,
    "args": {
      "Python parent id": 1442, "Python id": 1453, "finished": true, "Ev Idx": 8283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_50", "pid": 537, "tid": 537,
    "ts": 7429564830694.054, "dur": 46.505,
    "args": {
      "Python parent id": 1442, "Python module id": 50, "Python id": 1454, "finished": true, "Ev Idx": 8284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830694.746, "dur": 45.595,
    "args": {
      "Python parent id": 1454, "Python id": 1455, "finished": true, "Ev Idx": 8285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830694.972, "dur": 0.222,
    "args": {
      "Python parent id": 1455, "Python id": 1456, "finished": true, "Ev Idx": 8286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564830695.934, "dur": 44.266,
    "args": {
      "Python parent id": 1455, "Python id": 1457, "finished": true, "Ev Idx": 8287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830696.170, "dur": 0.410,
    "args": {
      "Python parent id": 1457, "Python id": 1458, "finished": true, "Ev Idx": 8288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830696.754, "dur": 0.118,
    "args": {
      "Python parent id": 1457, "Python id": 1459, "finished": true, "Ev Idx": 8289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564830696.976, "dur": 43.176,
    "args": {
      "Python parent id": 1457, "Python id": 1460, "finished": true, "Ev Idx": 8290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564830741.069, "dur": 2.385,
    "args": {
      "Python parent id": 1442, "Python id": 1461, "finished": true, "Ev Idx": 8291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564830743.890, "dur": 2.511,
    "args": {
      "Python parent id": 1442, "Python id": 1462, "finished": true, "Ev Idx": 8292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830747.565, "dur": 0.423,
    "args": {
      "Python parent id": 1442, "Python id": 1463, "finished": true, "Ev Idx": 8293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_51", "pid": 537, "tid": 537,
    "ts": 7429564830750.175, "dur": 44.744,
    "args": {
      "Python parent id": 1442, "Python module id": 51, "Python id": 1464, "finished": true, "Ev Idx": 8294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564830750.602, "dur": 44.131,
    "args": {
      "Python parent id": 1464, "Python id": 1465, "finished": true, "Ev Idx": 8295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564830750.851, "dur": 0.331,
    "args": {
      "Python parent id": 1465, "Python id": 1466, "finished": true, "Ev Idx": 8296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564830751.825, "dur": 42.764,
    "args": {
      "Python parent id": 1465, "Python id": 1467, "finished": true, "Ev Idx": 8297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830752.052, "dur": 0.142,
    "args": {
      "Python parent id": 1467, "Python id": 1468, "finished": true, "Ev Idx": 8298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564830752.384, "dur": 0.112,
    "args": {
      "Python parent id": 1467, "Python id": 1469, "finished": true, "Ev Idx": 8299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564830752.707, "dur": 41.834,
    "args": {
      "Python parent id": 1467, "Python id": 1470, "finished": true, "Ev Idx": 8300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564830795.482, "dur": 2.410,
    "args": {
      "Python parent id": 1442, "Python id": 1471, "finished": true, "Ev Idx": 8301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564830798.391, "dur": 2.694,
    "args": {
      "Python parent id": 1442, "Python id": 1472, "finished": true, "Ev Idx": 8302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564830802.427, "dur": 148.534,
    "args": {
      "Python parent id": 1442, "Python id": 1473, "finished": true, "Ev Idx": 8303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564830802.776, "dur": 3.654,
    "args": {
      "Python parent id": 1473, "Python id": 1474, "finished": true, "Ev Idx": 8304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564830806.754, "dur": 1.924,
    "args": {
      "Python parent id": 1473, "Python id": 1475, "finished": true, "Ev Idx": 8305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564830819.725, "dur": 51.541,
    "args": {
      "Python parent id": 1473, "Python id": 1476, "finished": true, "Ev Idx": 8306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564830847.801, "dur": 22.336,
    "args": {
      "Python parent id": 1476, "Python id": 1477, "finished": true, "Ev Idx": 8307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564830902.133, "dur": 28.442,
    "args": {
      "Python parent id": 1473, "Python id": 1478, "finished": true, "Ev Idx": 8308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564830917.889, "dur": 11.869,
    "args": {
      "Python parent id": 1478, "Python id": 1479, "finished": true, "Ev Idx": 8309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564830955.074, "dur": 4.054,
    "args": {
      "Python parent id": 1442, "Python id": 1480, "finished": true, "Ev Idx": 8310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564830955.796, "dur": 0.348,
    "args": {
      "Python parent id": 1480, "Python id": 1481, "finished": true, "Ev Idx": 8311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564830956.436, "dur": 0.093,
    "args": {
      "Python parent id": 1480, "Python id": 1482, "finished": true, "Ev Idx": 8312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564830958.041, "dur": 0.263,
    "args": {
      "Python parent id": 1480, "Python id": 1483, "finished": true, "Ev Idx": 8313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564830958.720, "dur": 0.164,
    "args": {
      "Python parent id": 1480, "Python id": 1484, "finished": true, "Ev Idx": 8314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564830960.422, "dur": 7.173,
    "args": {
      "Python parent id": 1442, "Python id": 1485, "finished": true, "Ev Idx": 8315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564830963.777, "dur": 3.679,
    "args": {
      "Python parent id": 1485, "Python id": 1486, "finished": true, "Ev Idx": 8316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564830964.078, "dur": 1.590,
    "args": {
      "Python parent id": 1486, "Python id": 1487, "finished": true, "Ev Idx": 8317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564830964.655, "dur": 0.861,
    "args": {
      "Python parent id": 1487, "Python id": 1488, "finished": true, "Ev Idx": 8318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564830965.937, "dur": 0.485,
    "args": {
      "Python parent id": 1486, "Python id": 1489, "finished": true, "Ev Idx": 8319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564830966.715, "dur": 0.669,
    "args": {
      "Python parent id": 1486, "Python id": 1490, "finished": true, "Ev Idx": 8320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564830970.262, "dur": 261.282,
    "args": {
      "Python parent id": 1442, "Python id": 1491, "finished": true, "Ev Idx": 8321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564830970.858, "dur": 47.157,
    "args": {
      "Python parent id": 1491, "Python id": 1492, "finished": true, "Ev Idx": 8322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564830989.407, "dur": 6.007,
    "args": {
      "Python parent id": 1492, "Python id": 1493, "finished": true, "Ev Idx": 8323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564830996.303, "dur": 21.587,
    "args": {
      "Python parent id": 1492, "Python id": 1494, "finished": true, "Ev Idx": 8324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564831018.846, "dur": 33.305,
    "args": {
      "Python parent id": 1491, "Python id": 1495, "finished": true, "Ev Idx": 8325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564831035.221, "dur": 2.645,
    "args": {
      "Python parent id": 1495, "Python id": 1496, "finished": true, "Ev Idx": 8326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564831038.351, "dur": 13.704,
    "args": {
      "Python parent id": 1495, "Python id": 1497, "finished": true, "Ev Idx": 8327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564831053.015, "dur": 3.776,
    "args": {
      "Python parent id": 1491, "Python id": 1498, "finished": true, "Ev Idx": 8328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564831057.162, "dur": 46.117,
    "args": {
      "Python parent id": 1491, "Python id": 1499, "finished": true, "Ev Idx": 8329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564831138.024, "dur": 16.405,
    "args": {
      "Python parent id": 1491, "Python id": 1500, "finished": true, "Ev Idx": 8330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564831138.263, "dur": 0.235,
    "args": {
      "Python parent id": 1500, "Python id": 1501, "finished": true, "Ev Idx": 8331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564831139.246, "dur": 15.024,
    "args": {
      "Python parent id": 1500, "Python id": 1502, "finished": true, "Ev Idx": 8332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564831155.383, "dur": 1.531,
    "args": {
      "Python parent id": 1491, "Python id": 1503, "finished": true, "Ev Idx": 8333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564831158.398, "dur": 10.377,
    "args": {
      "Python parent id": 1491, "Python id": 1504, "finished": true, "Ev Idx": 8334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564831158.528, "dur": 0.161,
    "args": {
      "Python parent id": 1504, "Python id": 1505, "finished": true, "Ev Idx": 8335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831163.918, "dur": 1.392,
    "args": {
      "Python parent id": 1504, "Python id": 1506, "finished": true, "Ev Idx": 8336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564831164.865, "dur": 0.321,
    "args": {
      "Python parent id": 1506, "Python id": 1507, "finished": true, "Ev Idx": 8337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564831165.722, "dur": 2.964,
    "args": {
      "Python parent id": 1504, "Python id": 1508, "finished": true, "Ev Idx": 8338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564831169.501, "dur": 42.459,
    "args": {
      "Python parent id": 1491, "Python id": 1509, "finished": true, "Ev Idx": 8339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564831212.580, "dur": 3.036,
    "args": {
      "Python parent id": 1491, "Python id": 1510, "finished": true, "Ev Idx": 8340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564831216.139, "dur": 14.333,
    "args": {
      "Python parent id": 1491, "Python id": 1511, "finished": true, "Ev Idx": 8341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564831234.613, "dur": 3.102,
    "args": {
      "Python parent id": 1442, "Python id": 1512, "finished": true, "Ev Idx": 8342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564831238.443, "dur": 0.261,
    "args": {
      "Python parent id": 1442, "Python id": 1513, "finished": true, "Ev Idx": 8343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831239.718, "dur": 0.922,
    "args": {
      "Python parent id": 1442, "Python id": 1514, "finished": true, "Ev Idx": 8344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_52", "pid": 537, "tid": 537,
    "ts": 7429564831245.104, "dur": 42.901,
    "args": {
      "Python parent id": 1442, "Python module id": 52, "Python id": 1515, "finished": true, "Ev Idx": 8345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831246.025, "dur": 41.553,
    "args": {
      "Python parent id": 1515, "Python id": 1516, "finished": true, "Ev Idx": 8346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831246.393, "dur": 1.102,
    "args": {
      "Python parent id": 1516, "Python id": 1517, "finished": true, "Ev Idx": 8347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564831248.682, "dur": 38.629,
    "args": {
      "Python parent id": 1516, "Python id": 1518, "finished": true, "Ev Idx": 8348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831249.333, "dur": 0.177,
    "args": {
      "Python parent id": 1518, "Python id": 1519, "finished": true, "Ev Idx": 8349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831249.796, "dur": 0.163,
    "args": {
      "Python parent id": 1518, "Python id": 1520, "finished": true, "Ev Idx": 8350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564831250.151, "dur": 37.055,
    "args": {
      "Python parent id": 1518, "Python id": 1521, "finished": true, "Ev Idx": 8351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831302.825, "dur": 0.683,
    "args": {
      "Python parent id": 1426, "Python id": 1522, "finished": true, "Ev Idx": 8352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_15", "pid": 537, "tid": 537,
    "ts": 7429564831306.320, "dur": 81.372,
    "args": {
      "Python parent id": 1426, "Python module id": 15, "Python id": 1523, "finished": true, "Ev Idx": 8353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831307.015, "dur": 80.369,
    "args": {
      "Python parent id": 1523, "Python id": 1524, "finished": true, "Ev Idx": 8354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831307.262, "dur": 0.405,
    "args": {
      "Python parent id": 1524, "Python id": 1525, "finished": true, "Ev Idx": 8355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564831308.675, "dur": 78.035,
    "args": {
      "Python parent id": 1524, "Python id": 1526, "finished": true, "Ev Idx": 8356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564831309.218, "dur": 1.725,
    "args": {
      "Python parent id": 1526, "Python id": 1527, "finished": true, "Ev Idx": 8357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564831311.268, "dur": 13.997,
    "args": {
      "Python parent id": 1526, "Python id": 1528, "finished": true, "Ev Idx": 8358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564831325.610, "dur": 14.301,
    "args": {
      "Python parent id": 1526, "Python id": 1529, "finished": true, "Ev Idx": 8359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564831354.008, "dur": 11.072,
    "args": {
      "Python parent id": 1526, "Python id": 1530, "finished": true, "Ev Idx": 8360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831375.278, "dur": 0.260,
    "args": {
      "Python parent id": 1526, "Python id": 1531, "finished": true, "Ev Idx": 8361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564831375.941, "dur": 1.341,
    "args": {
      "Python parent id": 1526, "Python id": 1532, "finished": true, "Ev Idx": 8362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831388.499, "dur": 0.391,
    "args": {
      "Python parent id": 1426, "Python id": 1533, "finished": true, "Ev Idx": 8363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_7", "pid": 537, "tid": 537,
    "ts": 7429564831390.881, "dur": 305.890,
    "args": {
      "Python parent id": 1426, "Python module id": 7, "Python id": 1534, "finished": true, "Ev Idx": 8364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831391.291, "dur": 305.276,
    "args": {
      "Python parent id": 1534, "Python id": 1535, "finished": true, "Ev Idx": 8365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831391.509, "dur": 0.306,
    "args": {
      "Python parent id": 1535, "Python id": 1536, "finished": true, "Ev Idx": 8366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564831392.725, "dur": 303.692,
    "args": {
      "Python parent id": 1535, "Python id": 1537, "finished": true, "Ev Idx": 8367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831393.019, "dur": 0.343,
    "args": {
      "Python parent id": 1537, "Python id": 1538, "finished": true, "Ev Idx": 8368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831393.736, "dur": 0.325,
    "args": {
      "Python parent id": 1537, "Python id": 1539, "finished": true, "Ev Idx": 8369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831394.367, "dur": 0.238,
    "args": {
      "Python parent id": 1537, "Python id": 1540, "finished": true, "Ev Idx": 8370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_53", "pid": 537, "tid": 537,
    "ts": 7429564831396.779, "dur": 128.404,
    "args": {
      "Python parent id": 1537, "Python module id": 53, "Python id": 1541, "finished": true, "Ev Idx": 8371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831397.406, "dur": 127.330,
    "args": {
      "Python parent id": 1541, "Python id": 1542, "finished": true, "Ev Idx": 8372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831397.548, "dur": 0.172,
    "args": {
      "Python parent id": 1542, "Python id": 1543, "finished": true, "Ev Idx": 8373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564831398.410, "dur": 126.007,
    "args": {
      "Python parent id": 1542, "Python id": 1544, "finished": true, "Ev Idx": 8374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831398.687, "dur": 0.154,
    "args": {
      "Python parent id": 1544, "Python id": 1545, "finished": true, "Ev Idx": 8375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831399.059, "dur": 0.289,
    "args": {
      "Python parent id": 1544, "Python id": 1546, "finished": true, "Ev Idx": 8376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564831399.484, "dur": 124.841,
    "args": {
      "Python parent id": 1544, "Python id": 1547, "finished": true, "Ev Idx": 8377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_7", "pid": 537, "tid": 537,
    "ts": 7429564831527.667, "dur": 19.773,
    "args": {
      "Python parent id": 1537, "Python module id": 7, "Python id": 1548, "finished": true, "Ev Idx": 8378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831528.928, "dur": 18.232,
    "args": {
      "Python parent id": 1548, "Python id": 1549, "finished": true, "Ev Idx": 8379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831529.187, "dur": 0.558,
    "args": {
      "Python parent id": 1549, "Python id": 1550, "finished": true, "Ev Idx": 8380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564831531.053, "dur": 15.963,
    "args": {
      "Python parent id": 1549, "Python id": 1551, "finished": true, "Ev Idx": 8381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564831532.021, "dur": 14.849,
    "args": {
      "Python parent id": 1551, "Python id": 1552, "finished": true, "Ev Idx": 8382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564831532.149, "dur": 0.237,
    "args": {
      "Python parent id": 1552, "Python id": 1553, "finished": true, "Ev Idx": 8383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564831533.291, "dur": 13.476,
    "args": {
      "Python parent id": 1552, "Python id": 1554, "finished": true, "Ev Idx": 8384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831548.150, "dur": 0.470,
    "args": {
      "Python parent id": 1537, "Python id": 1555, "finished": true, "Ev Idx": 8385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_54", "pid": 537, "tid": 537,
    "ts": 7429564831551.224, "dur": 84.463,
    "args": {
      "Python parent id": 1537, "Python module id": 54, "Python id": 1556, "finished": true, "Ev Idx": 8386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831551.873, "dur": 83.370,
    "args": {
      "Python parent id": 1556, "Python id": 1557, "finished": true, "Ev Idx": 8387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831552.095, "dur": 0.274,
    "args": {
      "Python parent id": 1557, "Python id": 1558, "finished": true, "Ev Idx": 8388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564831553.170, "dur": 81.818,
    "args": {
      "Python parent id": 1557, "Python id": 1559, "finished": true, "Ev Idx": 8389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831553.513, "dur": 0.148,
    "args": {
      "Python parent id": 1559, "Python id": 1560, "finished": true, "Ev Idx": 8390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831553.963, "dur": 0.119,
    "args": {
      "Python parent id": 1559, "Python id": 1561, "finished": true, "Ev Idx": 8391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564831554.287, "dur": 80.543,
    "args": {
      "Python parent id": 1559, "Python id": 1562, "finished": true, "Ev Idx": 8392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_55", "pid": 537, "tid": 537,
    "ts": 7429564831649.767, "dur": 46.288,
    "args": {
      "Python parent id": 1537, "Python module id": 55, "Python id": 1563, "finished": true, "Ev Idx": 8393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831650.530, "dur": 45.246,
    "args": {
      "Python parent id": 1563, "Python id": 1564, "finished": true, "Ev Idx": 8394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831650.791, "dur": 0.470,
    "args": {
      "Python parent id": 1564, "Python id": 1565, "finished": true, "Ev Idx": 8395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564831652.039, "dur": 43.643,
    "args": {
      "Python parent id": 1564, "Python id": 1566, "finished": true, "Ev Idx": 8396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831652.390, "dur": 0.211,
    "args": {
      "Python parent id": 1566, "Python id": 1567, "finished": true, "Ev Idx": 8397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831652.802, "dur": 0.153,
    "args": {
      "Python parent id": 1566, "Python id": 1568, "finished": true, "Ev Idx": 8398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564831653.070, "dur": 42.512,
    "args": {
      "Python parent id": 1566, "Python id": 1569, "finished": true, "Ev Idx": 8399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_8", "pid": 537, "tid": 537,
    "ts": 7429564831717.599, "dur": 1191.844,
    "args": {
      "Python parent id": 64, "Python module id": 8, "Python id": 1570, "finished": true, "Ev Idx": 8400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831719.261, "dur": 1189.507,
    "args": {
      "Python parent id": 1570, "Python id": 1571, "finished": true, "Ev Idx": 8401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831719.500, "dur": 0.430,
    "args": {
      "Python parent id": 1571, "Python id": 1572, "finished": true, "Ev Idx": 8402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564831721.761, "dur": 1186.149,
    "args": {
      "Python parent id": 1571, "Python id": 1573, "finished": true, "Ev Idx": 8403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831722.302, "dur": 0.606,
    "args": {
      "Python parent id": 1573, "Python id": 1574, "finished": true, "Ev Idx": 8404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_16", "pid": 537, "tid": 537,
    "ts": 7429564831724.947, "dur": 80.691,
    "args": {
      "Python parent id": 1573, "Python module id": 16, "Python id": 1575, "finished": true, "Ev Idx": 8405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831725.512, "dur": 79.878,
    "args": {
      "Python parent id": 1575, "Python id": 1576, "finished": true, "Ev Idx": 8406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831725.697, "dur": 0.163,
    "args": {
      "Python parent id": 1576, "Python id": 1577, "finished": true, "Ev Idx": 8407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564831726.903, "dur": 77.841,
    "args": {
      "Python parent id": 1576, "Python id": 1578, "finished": true, "Ev Idx": 8408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564831727.418, "dur": 2.064,
    "args": {
      "Python parent id": 1578, "Python id": 1579, "finished": true, "Ev Idx": 8409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564831729.796, "dur": 13.546,
    "args": {
      "Python parent id": 1578, "Python id": 1580, "finished": true, "Ev Idx": 8410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564831743.701, "dur": 15.150,
    "args": {
      "Python parent id": 1578, "Python id": 1581, "finished": true, "Ev Idx": 8411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564831771.302, "dur": 9.697,
    "args": {
      "Python parent id": 1578, "Python id": 1582, "finished": true, "Ev Idx": 8412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831792.407, "dur": 0.260,
    "args": {
      "Python parent id": 1578, "Python id": 1583, "finished": true, "Ev Idx": 8413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564831793.200, "dur": 1.176,
    "args": {
      "Python parent id": 1578, "Python id": 1584, "finished": true, "Ev Idx": 8414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831806.338, "dur": 0.377,
    "args": {
      "Python parent id": 1573, "Python id": 1585, "finished": true, "Ev Idx": 8415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_8", "pid": 537, "tid": 537,
    "ts": 7429564831809.849, "dur": 704.995,
    "args": {
      "Python parent id": 1573, "Python module id": 8, "Python id": 1586, "finished": true, "Ev Idx": 8416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831811.179, "dur": 703.192,
    "args": {
      "Python parent id": 1586, "Python id": 1587, "finished": true, "Ev Idx": 8417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831811.374, "dur": 0.282,
    "args": {
      "Python parent id": 1587, "Python id": 1588, "finished": true, "Ev Idx": 8418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564831813.335, "dur": 700.364,
    "args": {
      "Python parent id": 1587, "Python id": 1589, "finished": true, "Ev Idx": 8419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831816.856, "dur": 0.324,
    "args": {
      "Python parent id": 1589, "Python id": 1590, "finished": true, "Ev Idx": 8420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_56", "pid": 537, "tid": 537,
    "ts": 7429564831819.295, "dur": 43.224,
    "args": {
      "Python parent id": 1589, "Python module id": 56, "Python id": 1591, "finished": true, "Ev Idx": 8421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831819.887, "dur": 42.385,
    "args": {
      "Python parent id": 1591, "Python id": 1592, "finished": true, "Ev Idx": 8422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831820.038, "dur": 0.158,
    "args": {
      "Python parent id": 1592, "Python id": 1593, "finished": true, "Ev Idx": 8423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564831820.945, "dur": 41.185,
    "args": {
      "Python parent id": 1592, "Python id": 1594, "finished": true, "Ev Idx": 8424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831821.276, "dur": 0.139,
    "args": {
      "Python parent id": 1594, "Python id": 1595, "finished": true, "Ev Idx": 8425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831821.635, "dur": 0.112,
    "args": {
      "Python parent id": 1594, "Python id": 1596, "finished": true, "Ev Idx": 8426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564831821.884, "dur": 40.105,
    "args": {
      "Python parent id": 1594, "Python id": 1597, "finished": true, "Ev Idx": 8427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564831863.105, "dur": 3.816,
    "args": {
      "Python parent id": 1589, "Python id": 1598, "finished": true, "Ev Idx": 8428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564831867.398, "dur": 4.083,
    "args": {
      "Python parent id": 1589, "Python id": 1599, "finished": true, "Ev Idx": 8429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831872.471, "dur": 0.313,
    "args": {
      "Python parent id": 1589, "Python id": 1600, "finished": true, "Ev Idx": 8430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_57", "pid": 537, "tid": 537,
    "ts": 7429564831875.112, "dur": 46.441,
    "args": {
      "Python parent id": 1589, "Python module id": 57, "Python id": 1601, "finished": true, "Ev Idx": 8431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831875.695, "dur": 45.600,
    "args": {
      "Python parent id": 1601, "Python id": 1602, "finished": true, "Ev Idx": 8432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831875.928, "dur": 0.243,
    "args": {
      "Python parent id": 1602, "Python id": 1603, "finished": true, "Ev Idx": 8433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564831877.011, "dur": 44.157,
    "args": {
      "Python parent id": 1602, "Python id": 1604, "finished": true, "Ev Idx": 8434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831877.251, "dur": 0.172,
    "args": {
      "Python parent id": 1604, "Python id": 1605, "finished": true, "Ev Idx": 8435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831877.606, "dur": 0.156,
    "args": {
      "Python parent id": 1604, "Python id": 1606, "finished": true, "Ev Idx": 8436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564831877.885, "dur": 43.233,
    "args": {
      "Python parent id": 1604, "Python id": 1607, "finished": true, "Ev Idx": 8437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564831922.130, "dur": 2.521,
    "args": {
      "Python parent id": 1589, "Python id": 1608, "finished": true, "Ev Idx": 8438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564831925.395, "dur": 2.575,
    "args": {
      "Python parent id": 1589, "Python id": 1609, "finished": true, "Ev Idx": 8439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831928.655, "dur": 0.476,
    "args": {
      "Python parent id": 1589, "Python id": 1610, "finished": true, "Ev Idx": 8440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_58", "pid": 537, "tid": 537,
    "ts": 7429564831931.232, "dur": 43.283,
    "args": {
      "Python parent id": 1589, "Python module id": 58, "Python id": 1611, "finished": true, "Ev Idx": 8441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564831931.735, "dur": 42.558,
    "args": {
      "Python parent id": 1611, "Python id": 1612, "finished": true, "Ev Idx": 8442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564831931.913, "dur": 0.305,
    "args": {
      "Python parent id": 1612, "Python id": 1613, "finished": true, "Ev Idx": 8443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564831932.888, "dur": 41.216,
    "args": {
      "Python parent id": 1612, "Python id": 1614, "finished": true, "Ev Idx": 8444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831933.120, "dur": 0.152,
    "args": {
      "Python parent id": 1614, "Python id": 1615, "finished": true, "Ev Idx": 8445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564831933.472, "dur": 0.117,
    "args": {
      "Python parent id": 1614, "Python id": 1616, "finished": true, "Ev Idx": 8446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564831933.707, "dur": 40.304,
    "args": {
      "Python parent id": 1614, "Python id": 1617, "finished": true, "Ev Idx": 8447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564831975.132, "dur": 2.445,
    "args": {
      "Python parent id": 1589, "Python id": 1618, "finished": true, "Ev Idx": 8448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564831978.182, "dur": 2.656,
    "args": {
      "Python parent id": 1589, "Python id": 1619, "finished": true, "Ev Idx": 8449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564831981.972, "dur": 194.465,
    "args": {
      "Python parent id": 1589, "Python id": 1620, "finished": true, "Ev Idx": 8450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564831982.350, "dur": 3.505,
    "args": {
      "Python parent id": 1620, "Python id": 1621, "finished": true, "Ev Idx": 8451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564831986.171, "dur": 2.122,
    "args": {
      "Python parent id": 1620, "Python id": 1622, "finished": true, "Ev Idx": 8452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564832040.000, "dur": 54.384,
    "args": {
      "Python parent id": 1620, "Python id": 1623, "finished": true, "Ev Idx": 8453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564832071.242, "dur": 21.507,
    "args": {
      "Python parent id": 1623, "Python id": 1624, "finished": true, "Ev Idx": 8454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564832125.040, "dur": 33.094,
    "args": {
      "Python parent id": 1620, "Python id": 1625, "finished": true, "Ev Idx": 8455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564832143.357, "dur": 13.785,
    "args": {
      "Python parent id": 1625, "Python id": 1626, "finished": true, "Ev Idx": 8456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564832180.935, "dur": 4.250,
    "args": {
      "Python parent id": 1589, "Python id": 1627, "finished": true, "Ev Idx": 8457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564832181.793, "dur": 0.292,
    "args": {
      "Python parent id": 1627, "Python id": 1628, "finished": true, "Ev Idx": 8458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564832182.441, "dur": 0.064,
    "args": {
      "Python parent id": 1627, "Python id": 1629, "finished": true, "Ev Idx": 8459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564832183.975, "dur": 0.378,
    "args": {
      "Python parent id": 1627, "Python id": 1630, "finished": true, "Ev Idx": 8460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564832184.767, "dur": 0.158,
    "args": {
      "Python parent id": 1627, "Python id": 1631, "finished": true, "Ev Idx": 8461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564832186.366, "dur": 7.158,
    "args": {
      "Python parent id": 1589, "Python id": 1632, "finished": true, "Ev Idx": 8462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564832189.802, "dur": 3.559,
    "args": {
      "Python parent id": 1632, "Python id": 1633, "finished": true, "Ev Idx": 8463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564832190.044, "dur": 1.669,
    "args": {
      "Python parent id": 1633, "Python id": 1634, "finished": true, "Ev Idx": 8464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564832190.592, "dur": 0.972,
    "args": {
      "Python parent id": 1634, "Python id": 1635, "finished": true, "Ev Idx": 8465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564832192.006, "dur": 0.535,
    "args": {
      "Python parent id": 1633, "Python id": 1636, "finished": true, "Ev Idx": 8466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564832192.817, "dur": 0.464,
    "args": {
      "Python parent id": 1633, "Python id": 1637, "finished": true, "Ev Idx": 8467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564832196.378, "dur": 260.589,
    "args": {
      "Python parent id": 1589, "Python id": 1638, "finished": true, "Ev Idx": 8468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564832196.912, "dur": 47.505,
    "args": {
      "Python parent id": 1638, "Python id": 1639, "finished": true, "Ev Idx": 8469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564832216.335, "dur": 5.126,
    "args": {
      "Python parent id": 1639, "Python id": 1640, "finished": true, "Ev Idx": 8470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564832222.366, "dur": 21.904,
    "args": {
      "Python parent id": 1639, "Python id": 1641, "finished": true, "Ev Idx": 8471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564832245.301, "dur": 30.804,
    "args": {
      "Python parent id": 1638, "Python id": 1642, "finished": true, "Ev Idx": 8472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564832259.522, "dur": 3.404,
    "args": {
      "Python parent id": 1642, "Python id": 1643, "finished": true, "Ev Idx": 8473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564832263.419, "dur": 12.589,
    "args": {
      "Python parent id": 1642, "Python id": 1644, "finished": true, "Ev Idx": 8474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564832277.000, "dur": 3.064,
    "args": {
      "Python parent id": 1638, "Python id": 1645, "finished": true, "Ev Idx": 8475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564832280.486, "dur": 48.893,
    "args": {
      "Python parent id": 1638, "Python id": 1646, "finished": true, "Ev Idx": 8476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564832363.594, "dur": 17.195,
    "args": {
      "Python parent id": 1638, "Python id": 1647, "finished": true, "Ev Idx": 8477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564832363.764, "dur": 0.255,
    "args": {
      "Python parent id": 1647, "Python id": 1648, "finished": true, "Ev Idx": 8478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564832364.691, "dur": 15.976,
    "args": {
      "Python parent id": 1647, "Python id": 1649, "finished": true, "Ev Idx": 8479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564832381.708, "dur": 1.892,
    "args": {
      "Python parent id": 1638, "Python id": 1650, "finished": true, "Ev Idx": 8480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564832384.976, "dur": 9.460,
    "args": {
      "Python parent id": 1638, "Python id": 1651, "finished": true, "Ev Idx": 8481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564832385.048, "dur": 0.130,
    "args": {
      "Python parent id": 1651, "Python id": 1652, "finished": true, "Ev Idx": 8482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832390.141, "dur": 1.212,
    "args": {
      "Python parent id": 1651, "Python id": 1653, "finished": true, "Ev Idx": 8483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564832390.917, "dur": 0.372,
    "args": {
      "Python parent id": 1653, "Python id": 1654, "finished": true, "Ev Idx": 8484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564832391.710, "dur": 2.646,
    "args": {
      "Python parent id": 1651, "Python id": 1655, "finished": true, "Ev Idx": 8485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564832395.069, "dur": 42.895,
    "args": {
      "Python parent id": 1638, "Python id": 1656, "finished": true, "Ev Idx": 8486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564832438.590, "dur": 3.006,
    "args": {
      "Python parent id": 1638, "Python id": 1657, "finished": true, "Ev Idx": 8487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564832442.089, "dur": 13.879,
    "args": {
      "Python parent id": 1638, "Python id": 1658, "finished": true, "Ev Idx": 8488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564832459.581, "dur": 3.461,
    "args": {
      "Python parent id": 1589, "Python id": 1659, "finished": true, "Ev Idx": 8489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564832463.728, "dur": 0.301,
    "args": {
      "Python parent id": 1589, "Python id": 1660, "finished": true, "Ev Idx": 8490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832465.098, "dur": 0.930,
    "args": {
      "Python parent id": 1589, "Python id": 1661, "finished": true, "Ev Idx": 8491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_59", "pid": 537, "tid": 537,
    "ts": 7429564832470.457, "dur": 42.693,
    "args": {
      "Python parent id": 1589, "Python module id": 59, "Python id": 1662, "finished": true, "Ev Idx": 8492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564832471.353, "dur": 41.415,
    "args": {
      "Python parent id": 1662, "Python id": 1663, "finished": true, "Ev Idx": 8493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564832471.814, "dur": 1.325,
    "args": {
      "Python parent id": 1663, "Python id": 1664, "finished": true, "Ev Idx": 8494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564832474.208, "dur": 38.312,
    "args": {
      "Python parent id": 1663, "Python id": 1665, "finished": true, "Ev Idx": 8495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832474.608, "dur": 0.265,
    "args": {
      "Python parent id": 1665, "Python id": 1666, "finished": true, "Ev Idx": 8496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832475.129, "dur": 0.177,
    "args": {
      "Python parent id": 1665, "Python id": 1667, "finished": true, "Ev Idx": 8497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564832475.526, "dur": 36.897,
    "args": {
      "Python parent id": 1665, "Python id": 1668, "finished": true, "Ev Idx": 8498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832527.561, "dur": 0.686,
    "args": {
      "Python parent id": 1573, "Python id": 1669, "finished": true, "Ev Idx": 8499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_17", "pid": 537, "tid": 537,
    "ts": 7429564832530.908, "dur": 84.761,
    "args": {
      "Python parent id": 1573, "Python module id": 17, "Python id": 1670, "finished": true, "Ev Idx": 8500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564832531.463, "dur": 83.921,
    "args": {
      "Python parent id": 1670, "Python id": 1671, "finished": true, "Ev Idx": 8501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564832531.633, "dur": 0.387,
    "args": {
      "Python parent id": 1671, "Python id": 1672, "finished": true, "Ev Idx": 8502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564832533.148, "dur": 81.505,
    "args": {
      "Python parent id": 1671, "Python id": 1673, "finished": true, "Ev Idx": 8503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564832533.640, "dur": 1.674,
    "args": {
      "Python parent id": 1673, "Python id": 1674, "finished": true, "Ev Idx": 8504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564832535.601, "dur": 13.158,
    "args": {
      "Python parent id": 1673, "Python id": 1675, "finished": true, "Ev Idx": 8505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564832549.137, "dur": 16.519,
    "args": {
      "Python parent id": 1673, "Python id": 1676, "finished": true, "Ev Idx": 8506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564832578.279, "dur": 12.994,
    "args": {
      "Python parent id": 1673, "Python id": 1677, "finished": true, "Ev Idx": 8507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832601.374, "dur": 0.301,
    "args": {
      "Python parent id": 1673, "Python id": 1678, "finished": true, "Ev Idx": 8508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564832602.075, "dur": 1.193,
    "args": {
      "Python parent id": 1673, "Python id": 1679, "finished": true, "Ev Idx": 8509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832616.333, "dur": 0.361,
    "args": {
      "Python parent id": 1573, "Python id": 1680, "finished": true, "Ev Idx": 8510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_8", "pid": 537, "tid": 537,
    "ts": 7429564832618.561, "dur": 278.109,
    "args": {
      "Python parent id": 1573, "Python module id": 8, "Python id": 1681, "finished": true, "Ev Idx": 8511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564832618.977, "dur": 277.508,
    "args": {
      "Python parent id": 1681, "Python id": 1682, "finished": true, "Ev Idx": 8512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564832619.169, "dur": 0.334,
    "args": {
      "Python parent id": 1682, "Python id": 1683, "finished": true, "Ev Idx": 8513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564832620.555, "dur": 275.783,
    "args": {
      "Python parent id": 1682, "Python id": 1684, "finished": true, "Ev Idx": 8514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832620.852, "dur": 0.323,
    "args": {
      "Python parent id": 1684, "Python id": 1685, "finished": true, "Ev Idx": 8515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832621.568, "dur": 0.296,
    "args": {
      "Python parent id": 1684, "Python id": 1686, "finished": true, "Ev Idx": 8516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832622.130, "dur": 0.266,
    "args": {
      "Python parent id": 1684, "Python id": 1687, "finished": true, "Ev Idx": 8517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_60", "pid": 537, "tid": 537,
    "ts": 7429564832624.935, "dur": 54.405,
    "args": {
      "Python parent id": 1684, "Python module id": 60, "Python id": 1688, "finished": true, "Ev Idx": 8518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564832625.549, "dur": 53.474,
    "args": {
      "Python parent id": 1688, "Python id": 1689, "finished": true, "Ev Idx": 8519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564832625.692, "dur": 0.157,
    "args": {
      "Python parent id": 1689, "Python id": 1690, "finished": true, "Ev Idx": 8520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564832626.499, "dur": 52.397,
    "args": {
      "Python parent id": 1689, "Python id": 1691, "finished": true, "Ev Idx": 8521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832626.790, "dur": 0.137,
    "args": {
      "Python parent id": 1691, "Python id": 1692, "finished": true, "Ev Idx": 8522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832627.275, "dur": 0.112,
    "args": {
      "Python parent id": 1691, "Python id": 1693, "finished": true, "Ev Idx": 8523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564832627.543, "dur": 51.242,
    "args": {
      "Python parent id": 1691, "Python id": 1694, "finished": true, "Ev Idx": 8524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_8", "pid": 537, "tid": 537,
    "ts": 7429564832681.291, "dur": 69.082,
    "args": {
      "Python parent id": 1684, "Python module id": 8, "Python id": 1695, "finished": true, "Ev Idx": 8525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564832681.902, "dur": 68.135,
    "args": {
      "Python parent id": 1695, "Python id": 1696, "finished": true, "Ev Idx": 8526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564832682.078, "dur": 0.330,
    "args": {
      "Python parent id": 1696, "Python id": 1697, "finished": true, "Ev Idx": 8527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564832683.535, "dur": 66.353,
    "args": {
      "Python parent id": 1696, "Python id": 1698, "finished": true, "Ev Idx": 8528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564832684.548, "dur": 65.199,
    "args": {
      "Python parent id": 1698, "Python id": 1699, "finished": true, "Ev Idx": 8529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564832684.672, "dur": 0.230,
    "args": {
      "Python parent id": 1699, "Python id": 1700, "finished": true, "Ev Idx": 8530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564832685.751, "dur": 63.889,
    "args": {
      "Python parent id": 1699, "Python id": 1701, "finished": true, "Ev Idx": 8531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832751.162, "dur": 0.488,
    "args": {
      "Python parent id": 1684, "Python id": 1702, "finished": true, "Ev Idx": 8532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_61", "pid": 537, "tid": 537,
    "ts": 7429564832754.043, "dur": 50.695,
    "args": {
      "Python parent id": 1684, "Python module id": 61, "Python id": 1703, "finished": true, "Ev Idx": 8533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564832754.575, "dur": 49.874,
    "args": {
      "Python parent id": 1703, "Python id": 1704, "finished": true, "Ev Idx": 8534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564832754.767, "dur": 0.363,
    "args": {
      "Python parent id": 1704, "Python id": 1705, "finished": true, "Ev Idx": 8535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564832755.818, "dur": 48.520,
    "args": {
      "Python parent id": 1704, "Python id": 1706, "finished": true, "Ev Idx": 8536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832756.099, "dur": 0.116,
    "args": {
      "Python parent id": 1706, "Python id": 1707, "finished": true, "Ev Idx": 8537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832756.401, "dur": 0.119,
    "args": {
      "Python parent id": 1706, "Python id": 1708, "finished": true, "Ev Idx": 8538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564832756.623, "dur": 47.610,
    "args": {
      "Python parent id": 1706, "Python id": 1709, "finished": true, "Ev Idx": 8539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_62", "pid": 537, "tid": 537,
    "ts": 7429564832848.468, "dur": 47.550,
    "args": {
      "Python parent id": 1684, "Python module id": 62, "Python id": 1710, "finished": true, "Ev Idx": 8540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564832849.120, "dur": 46.638,
    "args": {
      "Python parent id": 1710, "Python id": 1711, "finished": true, "Ev Idx": 8541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564832849.351, "dur": 0.418,
    "args": {
      "Python parent id": 1711, "Python id": 1712, "finished": true, "Ev Idx": 8542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564832850.520, "dur": 45.093,
    "args": {
      "Python parent id": 1711, "Python id": 1713, "finished": true, "Ev Idx": 8543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832850.806, "dur": 0.195,
    "args": {
      "Python parent id": 1713, "Python id": 1714, "finished": true, "Ev Idx": 8544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832851.192, "dur": 0.144,
    "args": {
      "Python parent id": 1713, "Python id": 1715, "finished": true, "Ev Idx": 8545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564832851.571, "dur": 43.985,
    "args": {
      "Python parent id": 1713, "Python id": 1716, "finished": true, "Ev Idx": 8546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_9", "pid": 537, "tid": 537,
    "ts": 7429564832915.618, "dur": 1450.001,
    "args": {
      "Python parent id": 64, "Python module id": 9, "Python id": 1717, "finished": true, "Ev Idx": 8547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564832916.978, "dur": 1447.866,
    "args": {
      "Python parent id": 1717, "Python id": 1718, "finished": true, "Ev Idx": 8548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564832917.181, "dur": 0.400,
    "args": {
      "Python parent id": 1718, "Python id": 1719, "finished": true, "Ev Idx": 8549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564832919.096, "dur": 1444.480,
    "args": {
      "Python parent id": 1718, "Python id": 1720, "finished": true, "Ev Idx": 8550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832919.635, "dur": 0.604,
    "args": {
      "Python parent id": 1720, "Python id": 1721, "finished": true, "Ev Idx": 8551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_18", "pid": 537, "tid": 537,
    "ts": 7429564832922.425, "dur": 75.901,
    "args": {
      "Python parent id": 1720, "Python module id": 18, "Python id": 1722, "finished": true, "Ev Idx": 8552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564832922.980, "dur": 75.051,
    "args": {
      "Python parent id": 1722, "Python id": 1723, "finished": true, "Ev Idx": 8553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564832923.137, "dur": 0.164,
    "args": {
      "Python parent id": 1723, "Python id": 1724, "finished": true, "Ev Idx": 8554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564832924.139, "dur": 73.221,
    "args": {
      "Python parent id": 1723, "Python id": 1725, "finished": true, "Ev Idx": 8555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564832924.587, "dur": 1.912,
    "args": {
      "Python parent id": 1725, "Python id": 1726, "finished": true, "Ev Idx": 8556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564832926.729, "dur": 12.115,
    "args": {
      "Python parent id": 1725, "Python id": 1727, "finished": true, "Ev Idx": 8557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564832939.314, "dur": 13.958,
    "args": {
      "Python parent id": 1725, "Python id": 1728, "finished": true, "Ev Idx": 8558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564832967.007, "dur": 9.496,
    "args": {
      "Python parent id": 1725, "Python id": 1729, "finished": true, "Ev Idx": 8559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832987.104, "dur": 0.247,
    "args": {
      "Python parent id": 1725, "Python id": 1730, "finished": true, "Ev Idx": 8560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564832987.673, "dur": 1.178,
    "args": {
      "Python parent id": 1725, "Python id": 1731, "finished": true, "Ev Idx": 8561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564832998.981, "dur": 0.409,
    "args": {
      "Python parent id": 1720, "Python id": 1732, "finished": true, "Ev Idx": 8562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_9", "pid": 537, "tid": 537,
    "ts": 7429564833002.194, "dur": 869.620,
    "args": {
      "Python parent id": 1720, "Python module id": 9, "Python id": 1733, "finished": true, "Ev Idx": 8563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564833003.572, "dur": 867.597,
    "args": {
      "Python parent id": 1733, "Python id": 1734, "finished": true, "Ev Idx": 8564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564833003.734, "dur": 0.237,
    "args": {
      "Python parent id": 1734, "Python id": 1735, "finished": true, "Ev Idx": 8565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564833005.530, "dur": 864.614,
    "args": {
      "Python parent id": 1734, "Python id": 1736, "finished": true, "Ev Idx": 8566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833008.688, "dur": 0.286,
    "args": {
      "Python parent id": 1736, "Python id": 1737, "finished": true, "Ev Idx": 8567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_63", "pid": 537, "tid": 537,
    "ts": 7429564833011.471, "dur": 45.493,
    "args": {
      "Python parent id": 1736, "Python module id": 63, "Python id": 1738, "finished": true, "Ev Idx": 8568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564833012.074, "dur": 44.522,
    "args": {
      "Python parent id": 1738, "Python id": 1739, "finished": true, "Ev Idx": 8569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564833012.237, "dur": 0.168,
    "args": {
      "Python parent id": 1739, "Python id": 1740, "finished": true, "Ev Idx": 8570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564833013.107, "dur": 43.354,
    "args": {
      "Python parent id": 1739, "Python id": 1741, "finished": true, "Ev Idx": 8571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833013.421, "dur": 0.158,
    "args": {
      "Python parent id": 1741, "Python id": 1742, "finished": true, "Ev Idx": 8572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833013.789, "dur": 0.150,
    "args": {
      "Python parent id": 1741, "Python id": 1743, "finished": true, "Ev Idx": 8573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564833014.060, "dur": 42.313,
    "args": {
      "Python parent id": 1741, "Python id": 1744, "finished": true, "Ev Idx": 8574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564833057.805, "dur": 3.415,
    "args": {
      "Python parent id": 1736, "Python id": 1745, "finished": true, "Ev Idx": 8575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564833061.635, "dur": 3.857,
    "args": {
      "Python parent id": 1736, "Python id": 1746, "finished": true, "Ev Idx": 8576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833066.370, "dur": 0.405,
    "args": {
      "Python parent id": 1736, "Python id": 1747, "finished": true, "Ev Idx": 8577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_64", "pid": 537, "tid": 537,
    "ts": 7429564833068.791, "dur": 205.507,
    "args": {
      "Python parent id": 1736, "Python module id": 64, "Python id": 1748, "finished": true, "Ev Idx": 8578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564833069.213, "dur": 204.697,
    "args": {
      "Python parent id": 1748, "Python id": 1749, "finished": true, "Ev Idx": 8579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564833069.413, "dur": 0.268,
    "args": {
      "Python parent id": 1749, "Python id": 1750, "finished": true, "Ev Idx": 8580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564833070.578, "dur": 203.049,
    "args": {
      "Python parent id": 1749, "Python id": 1751, "finished": true, "Ev Idx": 8581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833220.738, "dur": 0.262,
    "args": {
      "Python parent id": 1751, "Python id": 1752, "finished": true, "Ev Idx": 8582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833221.308, "dur": 0.127,
    "args": {
      "Python parent id": 1751, "Python id": 1753, "finished": true, "Ev Idx": 8583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564833221.693, "dur": 51.867,
    "args": {
      "Python parent id": 1751, "Python id": 1754, "finished": true, "Ev Idx": 8584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564833275.006, "dur": 3.277,
    "args": {
      "Python parent id": 1736, "Python id": 1755, "finished": true, "Ev Idx": 8585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564833278.954, "dur": 3.682,
    "args": {
      "Python parent id": 1736, "Python id": 1756, "finished": true, "Ev Idx": 8586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833283.710, "dur": 0.414,
    "args": {
      "Python parent id": 1736, "Python id": 1757, "finished": true, "Ev Idx": 8587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_65", "pid": 537, "tid": 537,
    "ts": 7429564833286.921, "dur": 46.792,
    "args": {
      "Python parent id": 1736, "Python module id": 65, "Python id": 1758, "finished": true, "Ev Idx": 8588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564833287.653, "dur": 45.652,
    "args": {
      "Python parent id": 1758, "Python id": 1759, "finished": true, "Ev Idx": 8589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564833288.000, "dur": 0.465,
    "args": {
      "Python parent id": 1759, "Python id": 1760, "finished": true, "Ev Idx": 8590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564833289.230, "dur": 43.849,
    "args": {
      "Python parent id": 1759, "Python id": 1761, "finished": true, "Ev Idx": 8591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833289.478, "dur": 0.231,
    "args": {
      "Python parent id": 1761, "Python id": 1762, "finished": true, "Ev Idx": 8592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833289.897, "dur": 0.130,
    "args": {
      "Python parent id": 1761, "Python id": 1763, "finished": true, "Ev Idx": 8593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564833290.155, "dur": 42.839,
    "args": {
      "Python parent id": 1761, "Python id": 1764, "finished": true, "Ev Idx": 8594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564833334.309, "dur": 2.597,
    "args": {
      "Python parent id": 1736, "Python id": 1765, "finished": true, "Ev Idx": 8595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564833337.349, "dur": 2.843,
    "args": {
      "Python parent id": 1736, "Python id": 1766, "finished": true, "Ev Idx": 8596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564833341.554, "dur": 187.434,
    "args": {
      "Python parent id": 1736, "Python id": 1767, "finished": true, "Ev Idx": 8597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564833342.032, "dur": 34.578,
    "args": {
      "Python parent id": 1767, "Python id": 1768, "finished": true, "Ev Idx": 8598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564833376.950, "dur": 1.946,
    "args": {
      "Python parent id": 1767, "Python id": 1769, "finished": true, "Ev Idx": 8599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564833391.741, "dur": 55.483,
    "args": {
      "Python parent id": 1767, "Python id": 1770, "finished": true, "Ev Idx": 8600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564833422.934, "dur": 23.130,
    "args": {
      "Python parent id": 1770, "Python id": 1771, "finished": true, "Ev Idx": 8601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564833478.967, "dur": 29.075,
    "args": {
      "Python parent id": 1767, "Python id": 1772, "finished": true, "Ev Idx": 8602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564833495.748, "dur": 11.478,
    "args": {
      "Python parent id": 1772, "Python id": 1773, "finished": true, "Ev Idx": 8603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564833533.364, "dur": 4.152,
    "args": {
      "Python parent id": 1736, "Python id": 1774, "finished": true, "Ev Idx": 8604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564833534.278, "dur": 0.308,
    "args": {
      "Python parent id": 1774, "Python id": 1775, "finished": true, "Ev Idx": 8605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564833535.059, "dur": 0.040,
    "args": {
      "Python parent id": 1774, "Python id": 1776, "finished": true, "Ev Idx": 8606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564833536.488, "dur": 0.260,
    "args": {
      "Python parent id": 1774, "Python id": 1777, "finished": true, "Ev Idx": 8607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564833537.149, "dur": 0.103,
    "args": {
      "Python parent id": 1774, "Python id": 1778, "finished": true, "Ev Idx": 8608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564833538.818, "dur": 7.218,
    "args": {
      "Python parent id": 1736, "Python id": 1779, "finished": true, "Ev Idx": 8609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564833542.315, "dur": 3.503,
    "args": {
      "Python parent id": 1779, "Python id": 1780, "finished": true, "Ev Idx": 8610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564833542.540, "dur": 1.608,
    "args": {
      "Python parent id": 1780, "Python id": 1781, "finished": true, "Ev Idx": 8611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564833543.208, "dur": 0.767,
    "args": {
      "Python parent id": 1781, "Python id": 1782, "finished": true, "Ev Idx": 8612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564833544.425, "dur": 0.523,
    "args": {
      "Python parent id": 1780, "Python id": 1783, "finished": true, "Ev Idx": 8613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564833545.283, "dur": 0.455,
    "args": {
      "Python parent id": 1780, "Python id": 1784, "finished": true, "Ev Idx": 8614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564833549.071, "dur": 264.116,
    "args": {
      "Python parent id": 1736, "Python id": 1785, "finished": true, "Ev Idx": 8615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564833549.775, "dur": 49.209,
    "args": {
      "Python parent id": 1785, "Python id": 1786, "finished": true, "Ev Idx": 8616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564833568.816, "dur": 5.878,
    "args": {
      "Python parent id": 1786, "Python id": 1787, "finished": true, "Ev Idx": 8617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564833575.583, "dur": 23.239,
    "args": {
      "Python parent id": 1786, "Python id": 1788, "finished": true, "Ev Idx": 8618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564833599.779, "dur": 32.645,
    "args": {
      "Python parent id": 1785, "Python id": 1789, "finished": true, "Ev Idx": 8619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564833615.384, "dur": 2.732,
    "args": {
      "Python parent id": 1789, "Python id": 1790, "finished": true, "Ev Idx": 8620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564833618.583, "dur": 13.773,
    "args": {
      "Python parent id": 1789, "Python id": 1791, "finished": true, "Ev Idx": 8621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564833633.466, "dur": 2.820,
    "args": {
      "Python parent id": 1785, "Python id": 1792, "finished": true, "Ev Idx": 8622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564833636.607, "dur": 48.631,
    "args": {
      "Python parent id": 1785, "Python id": 1793, "finished": true, "Ev Idx": 8623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564833718.902, "dur": 18.666,
    "args": {
      "Python parent id": 1785, "Python id": 1794, "finished": true, "Ev Idx": 8624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564833719.227, "dur": 0.260,
    "args": {
      "Python parent id": 1794, "Python id": 1795, "finished": true, "Ev Idx": 8625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564833720.186, "dur": 17.255,
    "args": {
      "Python parent id": 1794, "Python id": 1796, "finished": true, "Ev Idx": 8626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564833738.466, "dur": 1.965,
    "args": {
      "Python parent id": 1785, "Python id": 1797, "finished": true, "Ev Idx": 8627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564833742.094, "dur": 10.445,
    "args": {
      "Python parent id": 1785, "Python id": 1798, "finished": true, "Ev Idx": 8628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564833742.210, "dur": 0.068,
    "args": {
      "Python parent id": 1798, "Python id": 1799, "finished": true, "Ev Idx": 8629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833747.887, "dur": 1.320,
    "args": {
      "Python parent id": 1798, "Python id": 1800, "finished": true, "Ev Idx": 8630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564833748.645, "dur": 0.475,
    "args": {
      "Python parent id": 1800, "Python id": 1801, "finished": true, "Ev Idx": 8631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564833749.622, "dur": 2.837,
    "args": {
      "Python parent id": 1798, "Python id": 1802, "finished": true, "Ev Idx": 8632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564833753.190, "dur": 40.587,
    "args": {
      "Python parent id": 1785, "Python id": 1803, "finished": true, "Ev Idx": 8633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564833794.417, "dur": 3.183,
    "args": {
      "Python parent id": 1785, "Python id": 1804, "finished": true, "Ev Idx": 8634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564833798.347, "dur": 13.866,
    "args": {
      "Python parent id": 1785, "Python id": 1805, "finished": true, "Ev Idx": 8635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564833816.176, "dur": 3.197,
    "args": {
      "Python parent id": 1736, "Python id": 1806, "finished": true, "Ev Idx": 8636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564833820.098, "dur": 0.282,
    "args": {
      "Python parent id": 1736, "Python id": 1807, "finished": true, "Ev Idx": 8637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833821.450, "dur": 0.837,
    "args": {
      "Python parent id": 1736, "Python id": 1808, "finished": true, "Ev Idx": 8638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_66", "pid": 537, "tid": 537,
    "ts": 7429564833826.461, "dur": 43.055,
    "args": {
      "Python parent id": 1736, "Python module id": 66, "Python id": 1809, "finished": true, "Ev Idx": 8639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564833827.405, "dur": 41.622,
    "args": {
      "Python parent id": 1809, "Python id": 1810, "finished": true, "Ev Idx": 8640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564833827.778, "dur": 1.034,
    "args": {
      "Python parent id": 1810, "Python id": 1811, "finished": true, "Ev Idx": 8641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564833830.025, "dur": 38.768,
    "args": {
      "Python parent id": 1810, "Python id": 1812, "finished": true, "Ev Idx": 8642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833830.438, "dur": 0.238,
    "args": {
      "Python parent id": 1812, "Python id": 1813, "finished": true, "Ev Idx": 8643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833830.967, "dur": 0.148,
    "args": {
      "Python parent id": 1812, "Python id": 1814, "finished": true, "Ev Idx": 8644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564833831.320, "dur": 37.398,
    "args": {
      "Python parent id": 1812, "Python id": 1815, "finished": true, "Ev Idx": 8645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564833952.590, "dur": 0.835,
    "args": {
      "Python parent id": 1720, "Python id": 1816, "finished": true, "Ev Idx": 8646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_19", "pid": 537, "tid": 537,
    "ts": 7429564833969.514, "dur": 89.336,
    "args": {
      "Python parent id": 1720, "Python module id": 19, "Python id": 1817, "finished": true, "Ev Idx": 8647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564833970.374, "dur": 88.054,
    "args": {
      "Python parent id": 1817, "Python id": 1818, "finished": true, "Ev Idx": 8648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564833970.624, "dur": 0.844,
    "args": {
      "Python parent id": 1818, "Python id": 1819, "finished": true, "Ev Idx": 8649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564833972.885, "dur": 84.706,
    "args": {
      "Python parent id": 1818, "Python id": 1820, "finished": true, "Ev Idx": 8650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564833973.599, "dur": 2.262,
    "args": {
      "Python parent id": 1820, "Python id": 1821, "finished": true, "Ev Idx": 8651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564833976.166, "dur": 16.722,
    "args": {
      "Python parent id": 1820, "Python id": 1822, "finished": true, "Ev Idx": 8652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564833993.227, "dur": 17.563,
    "args": {
      "Python parent id": 1820, "Python id": 1823, "finished": true, "Ev Idx": 8653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564834023.818, "dur": 12.125,
    "args": {
      "Python parent id": 1820, "Python id": 1824, "finished": true, "Ev Idx": 8654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834047.265, "dur": 0.263,
    "args": {
      "Python parent id": 1820, "Python id": 1825, "finished": true, "Ev Idx": 8655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564834047.895, "dur": 1.391,
    "args": {
      "Python parent id": 1820, "Python id": 1826, "finished": true, "Ev Idx": 8656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834059.559, "dur": 0.409,
    "args": {
      "Python parent id": 1720, "Python id": 1827, "finished": true, "Ev Idx": 8657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_9", "pid": 537, "tid": 537,
    "ts": 7429564834061.783, "dur": 290.468,
    "args": {
      "Python parent id": 1720, "Python module id": 9, "Python id": 1828, "finished": true, "Ev Idx": 8658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834062.403, "dur": 289.612,
    "args": {
      "Python parent id": 1828, "Python id": 1829, "finished": true, "Ev Idx": 8659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834062.645, "dur": 0.414,
    "args": {
      "Python parent id": 1829, "Python id": 1830, "finished": true, "Ev Idx": 8660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564834064.218, "dur": 287.654,
    "args": {
      "Python parent id": 1829, "Python id": 1831, "finished": true, "Ev Idx": 8661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834064.537, "dur": 0.373,
    "args": {
      "Python parent id": 1831, "Python id": 1832, "finished": true, "Ev Idx": 8662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834065.243, "dur": 0.289,
    "args": {
      "Python parent id": 1831, "Python id": 1833, "finished": true, "Ev Idx": 8663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834065.801, "dur": 0.278,
    "args": {
      "Python parent id": 1831, "Python id": 1834, "finished": true, "Ev Idx": 8664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_67", "pid": 537, "tid": 537,
    "ts": 7429564834067.849, "dur": 94.260,
    "args": {
      "Python parent id": 1831, "Python module id": 67, "Python id": 1835, "finished": true, "Ev Idx": 8665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834068.458, "dur": 93.201,
    "args": {
      "Python parent id": 1835, "Python id": 1836, "finished": true, "Ev Idx": 8666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834068.600, "dur": 0.214,
    "args": {
      "Python parent id": 1836, "Python id": 1837, "finished": true, "Ev Idx": 8667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564834089.523, "dur": 71.890,
    "args": {
      "Python parent id": 1836, "Python id": 1838, "finished": true, "Ev Idx": 8668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834090.091, "dur": 0.314,
    "args": {
      "Python parent id": 1838, "Python id": 1839, "finished": true, "Ev Idx": 8669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834090.689, "dur": 3.416,
    "args": {
      "Python parent id": 1838, "Python id": 1840, "finished": true, "Ev Idx": 8670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564834094.481, "dur": 66.810,
    "args": {
      "Python parent id": 1838, "Python id": 1841, "finished": true, "Ev Idx": 8671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_9", "pid": 537, "tid": 537,
    "ts": 7429564834164.490, "dur": 19.035,
    "args": {
      "Python parent id": 1831, "Python module id": 9, "Python id": 1842, "finished": true, "Ev Idx": 8672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834165.217, "dur": 18.074,
    "args": {
      "Python parent id": 1842, "Python id": 1843, "finished": true, "Ev Idx": 8673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834165.496, "dur": 0.485,
    "args": {
      "Python parent id": 1843, "Python id": 1844, "finished": true, "Ev Idx": 8674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564834167.098, "dur": 16.041,
    "args": {
      "Python parent id": 1843, "Python id": 1845, "finished": true, "Ev Idx": 8675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564834168.120, "dur": 14.857,
    "args": {
      "Python parent id": 1845, "Python id": 1846, "finished": true, "Ev Idx": 8676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564834168.347, "dur": 0.279,
    "args": {
      "Python parent id": 1846, "Python id": 1847, "finished": true, "Ev Idx": 8677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564834169.594, "dur": 13.292,
    "args": {
      "Python parent id": 1846, "Python id": 1848, "finished": true, "Ev Idx": 8678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834184.180, "dur": 0.459,
    "args": {
      "Python parent id": 1831, "Python id": 1849, "finished": true, "Ev Idx": 8679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_68", "pid": 537, "tid": 537,
    "ts": 7429564834186.901, "dur": 100.650,
    "args": {
      "Python parent id": 1831, "Python module id": 68, "Python id": 1850, "finished": true, "Ev Idx": 8680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834187.503, "dur": 99.622,
    "args": {
      "Python parent id": 1850, "Python id": 1851, "finished": true, "Ev Idx": 8681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834187.664, "dur": 0.233,
    "args": {
      "Python parent id": 1851, "Python id": 1852, "finished": true, "Ev Idx": 8682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564834188.617, "dur": 98.303,
    "args": {
      "Python parent id": 1851, "Python id": 1853, "finished": true, "Ev Idx": 8683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834188.891, "dur": 0.170,
    "args": {
      "Python parent id": 1853, "Python id": 1854, "finished": true, "Ev Idx": 8684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834189.226, "dur": 0.174,
    "args": {
      "Python parent id": 1853, "Python id": 1855, "finished": true, "Ev Idx": 8685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564834189.519, "dur": 97.270,
    "args": {
      "Python parent id": 1853, "Python id": 1856, "finished": true, "Ev Idx": 8686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_69", "pid": 537, "tid": 537,
    "ts": 7429564834302.080, "dur": 49.357,
    "args": {
      "Python parent id": 1831, "Python module id": 69, "Python id": 1857, "finished": true, "Ev Idx": 8687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834303.006, "dur": 47.988,
    "args": {
      "Python parent id": 1857, "Python id": 1858, "finished": true, "Ev Idx": 8688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834303.295, "dur": 0.484,
    "args": {
      "Python parent id": 1858, "Python id": 1859, "finished": true, "Ev Idx": 8689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564834304.592, "dur": 46.285,
    "args": {
      "Python parent id": 1858, "Python id": 1860, "finished": true, "Ev Idx": 8690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834304.960, "dur": 0.289,
    "args": {
      "Python parent id": 1860, "Python id": 1861, "finished": true, "Ev Idx": 8691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834305.470, "dur": 0.115,
    "args": {
      "Python parent id": 1860, "Python id": 1862, "finished": true, "Ev Idx": 8692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564834305.692, "dur": 45.067,
    "args": {
      "Python parent id": 1860, "Python id": 1863, "finished": true, "Ev Idx": 8693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_10", "pid": 537, "tid": 537,
    "ts": 7429564834372.258, "dur": 1210.355,
    "args": {
      "Python parent id": 64, "Python module id": 10, "Python id": 1864, "finished": true, "Ev Idx": 8694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834373.703, "dur": 1208.171,
    "args": {
      "Python parent id": 1864, "Python id": 1865, "finished": true, "Ev Idx": 8695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834373.892, "dur": 0.360,
    "args": {
      "Python parent id": 1865, "Python id": 1866, "finished": true, "Ev Idx": 8696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564834376.154, "dur": 1204.752,
    "args": {
      "Python parent id": 1865, "Python id": 1867, "finished": true, "Ev Idx": 8697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834376.727, "dur": 0.591,
    "args": {
      "Python parent id": 1867, "Python id": 1868, "finished": true, "Ev Idx": 8698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_20", "pid": 537, "tid": 537,
    "ts": 7429564834379.652, "dur": 118.817,
    "args": {
      "Python parent id": 1867, "Python module id": 20, "Python id": 1869, "finished": true, "Ev Idx": 8699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834380.192, "dur": 117.769,
    "args": {
      "Python parent id": 1869, "Python id": 1870, "finished": true, "Ev Idx": 8700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834380.343, "dur": 0.173,
    "args": {
      "Python parent id": 1870, "Python id": 1871, "finished": true, "Ev Idx": 8701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564834381.459, "dur": 115.588,
    "args": {
      "Python parent id": 1870, "Python id": 1872, "finished": true, "Ev Idx": 8702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564834382.117, "dur": 2.066,
    "args": {
      "Python parent id": 1872, "Python id": 1873, "finished": true, "Ev Idx": 8703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564834384.487, "dur": 12.519,
    "args": {
      "Python parent id": 1872, "Python id": 1874, "finished": true, "Ev Idx": 8704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564834397.351, "dur": 15.024,
    "args": {
      "Python parent id": 1872, "Python id": 1875, "finished": true, "Ev Idx": 8705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564834425.561, "dur": 11.566,
    "args": {
      "Python parent id": 1872, "Python id": 1876, "finished": true, "Ev Idx": 8706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834448.266, "dur": 0.257,
    "args": {
      "Python parent id": 1872, "Python id": 1877, "finished": true, "Ev Idx": 8707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564834448.988, "dur": 1.395,
    "args": {
      "Python parent id": 1872, "Python id": 1878, "finished": true, "Ev Idx": 8708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834499.313, "dur": 0.436,
    "args": {
      "Python parent id": 1867, "Python id": 1879, "finished": true, "Ev Idx": 8709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_10", "pid": 537, "tid": 537,
    "ts": 7429564834502.900, "dur": 682.002,
    "args": {
      "Python parent id": 1867, "Python module id": 10, "Python id": 1880, "finished": true, "Ev Idx": 8710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834504.798, "dur": 679.474,
    "args": {
      "Python parent id": 1880, "Python id": 1881, "finished": true, "Ev Idx": 8711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834505.055, "dur": 0.437,
    "args": {
      "Python parent id": 1881, "Python id": 1882, "finished": true, "Ev Idx": 8712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564834507.172, "dur": 676.236,
    "args": {
      "Python parent id": 1881, "Python id": 1883, "finished": true, "Ev Idx": 8713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834510.984, "dur": 0.331,
    "args": {
      "Python parent id": 1883, "Python id": 1884, "finished": true, "Ev Idx": 8714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_70", "pid": 537, "tid": 537,
    "ts": 7429564834513.553, "dur": 52.421,
    "args": {
      "Python parent id": 1883, "Python module id": 70, "Python id": 1885, "finished": true, "Ev Idx": 8715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834514.231, "dur": 51.462,
    "args": {
      "Python parent id": 1885, "Python id": 1886, "finished": true, "Ev Idx": 8716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834514.386, "dur": 0.159,
    "args": {
      "Python parent id": 1886, "Python id": 1887, "finished": true, "Ev Idx": 8717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564834515.321, "dur": 50.236,
    "args": {
      "Python parent id": 1886, "Python id": 1888, "finished": true, "Ev Idx": 8718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834515.736, "dur": 0.162,
    "args": {
      "Python parent id": 1888, "Python id": 1889, "finished": true, "Ev Idx": 8719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834516.156, "dur": 0.145,
    "args": {
      "Python parent id": 1888, "Python id": 1890, "finished": true, "Ev Idx": 8720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564834516.468, "dur": 48.957,
    "args": {
      "Python parent id": 1888, "Python id": 1891, "finished": true, "Ev Idx": 8721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564834566.640, "dur": 4.182,
    "args": {
      "Python parent id": 1883, "Python id": 1892, "finished": true, "Ev Idx": 8722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564834571.321, "dur": 4.640,
    "args": {
      "Python parent id": 1883, "Python id": 1893, "finished": true, "Ev Idx": 8723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834576.902, "dur": 0.370,
    "args": {
      "Python parent id": 1883, "Python id": 1894, "finished": true, "Ev Idx": 8724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_71", "pid": 537, "tid": 537,
    "ts": 7429564834579.364, "dur": 48.777,
    "args": {
      "Python parent id": 1883, "Python module id": 71, "Python id": 1895, "finished": true, "Ev Idx": 8725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834579.892, "dur": 48.009,
    "args": {
      "Python parent id": 1895, "Python id": 1896, "finished": true, "Ev Idx": 8726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834580.164, "dur": 0.302,
    "args": {
      "Python parent id": 1896, "Python id": 1897, "finished": true, "Ev Idx": 8727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564834581.163, "dur": 46.593,
    "args": {
      "Python parent id": 1896, "Python id": 1898, "finished": true, "Ev Idx": 8728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834581.377, "dur": 0.177,
    "args": {
      "Python parent id": 1898, "Python id": 1899, "finished": true, "Ev Idx": 8729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834581.734, "dur": 0.122,
    "args": {
      "Python parent id": 1898, "Python id": 1900, "finished": true, "Ev Idx": 8730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564834582.010, "dur": 45.670,
    "args": {
      "Python parent id": 1898, "Python id": 1901, "finished": true, "Ev Idx": 8731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564834628.671, "dur": 2.499,
    "args": {
      "Python parent id": 1883, "Python id": 1902, "finished": true, "Ev Idx": 8732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564834631.896, "dur": 2.778,
    "args": {
      "Python parent id": 1883, "Python id": 1903, "finished": true, "Ev Idx": 8733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834635.669, "dur": 0.516,
    "args": {
      "Python parent id": 1883, "Python id": 1904, "finished": true, "Ev Idx": 8734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_72", "pid": 537, "tid": 537,
    "ts": 7429564834638.279, "dur": 41.979,
    "args": {
      "Python parent id": 1883, "Python module id": 72, "Python id": 1905, "finished": true, "Ev Idx": 8735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564834638.730, "dur": 41.244,
    "args": {
      "Python parent id": 1905, "Python id": 1906, "finished": true, "Ev Idx": 8736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564834638.989, "dur": 0.345,
    "args": {
      "Python parent id": 1906, "Python id": 1907, "finished": true, "Ev Idx": 8737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564834640.042, "dur": 39.841,
    "args": {
      "Python parent id": 1906, "Python id": 1908, "finished": true, "Ev Idx": 8738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834640.280, "dur": 0.154,
    "args": {
      "Python parent id": 1908, "Python id": 1909, "finished": true, "Ev Idx": 8739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564834640.627, "dur": 0.111,
    "args": {
      "Python parent id": 1908, "Python id": 1910, "finished": true, "Ev Idx": 8740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564834640.861, "dur": 38.925,
    "args": {
      "Python parent id": 1908, "Python id": 1911, "finished": true, "Ev Idx": 8741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564834680.770, "dur": 2.490,
    "args": {
      "Python parent id": 1883, "Python id": 1912, "finished": true, "Ev Idx": 8742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564834683.746, "dur": 2.413,
    "args": {
      "Python parent id": 1883, "Python id": 1913, "finished": true, "Ev Idx": 8743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564834687.535, "dur": 152.206,
    "args": {
      "Python parent id": 1883, "Python id": 1914, "finished": true, "Ev Idx": 8744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564834687.950, "dur": 3.729,
    "args": {
      "Python parent id": 1914, "Python id": 1915, "finished": true, "Ev Idx": 8745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564834692.018, "dur": 1.788,
    "args": {
      "Python parent id": 1914, "Python id": 1916, "finished": true, "Ev Idx": 8746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564834706.107, "dur": 52.807,
    "args": {
      "Python parent id": 1914, "Python id": 1917, "finished": true, "Ev Idx": 8747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564834734.817, "dur": 22.631,
    "args": {
      "Python parent id": 1917, "Python id": 1918, "finished": true, "Ev Idx": 8748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564834791.976, "dur": 29.347,
    "args": {
      "Python parent id": 1914, "Python id": 1919, "finished": true, "Ev Idx": 8749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564834809.289, "dur": 11.190,
    "args": {
      "Python parent id": 1919, "Python id": 1920, "finished": true, "Ev Idx": 8750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564834844.466, "dur": 4.280,
    "args": {
      "Python parent id": 1883, "Python id": 1921, "finished": true, "Ev Idx": 8751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564834845.373, "dur": 0.367,
    "args": {
      "Python parent id": 1921, "Python id": 1922, "finished": true, "Ev Idx": 8752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564834846.169, "dur": 0.036,
    "args": {
      "Python parent id": 1921, "Python id": 1923, "finished": true, "Ev Idx": 8753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564834847.631, "dur": 0.238,
    "args": {
      "Python parent id": 1921, "Python id": 1924, "finished": true, "Ev Idx": 8754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564834848.317, "dur": 0.157,
    "args": {
      "Python parent id": 1921, "Python id": 1925, "finished": true, "Ev Idx": 8755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564834850.337, "dur": 7.061,
    "args": {
      "Python parent id": 1883, "Python id": 1926, "finished": true, "Ev Idx": 8756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564834853.837, "dur": 3.386,
    "args": {
      "Python parent id": 1926, "Python id": 1927, "finished": true, "Ev Idx": 8757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564834854.059, "dur": 1.524,
    "args": {
      "Python parent id": 1927, "Python id": 1928, "finished": true, "Ev Idx": 8758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564834854.593, "dur": 0.809,
    "args": {
      "Python parent id": 1928, "Python id": 1929, "finished": true, "Ev Idx": 8759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564834855.893, "dur": 0.506,
    "args": {
      "Python parent id": 1927, "Python id": 1930, "finished": true, "Ev Idx": 8760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564834856.703, "dur": 0.446,
    "args": {
      "Python parent id": 1927, "Python id": 1931, "finished": true, "Ev Idx": 8761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564834860.345, "dur": 264.774,
    "args": {
      "Python parent id": 1883, "Python id": 1932, "finished": true, "Ev Idx": 8762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564834860.926, "dur": 48.201,
    "args": {
      "Python parent id": 1932, "Python id": 1933, "finished": true, "Ev Idx": 8763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564834879.870, "dur": 5.495,
    "args": {
      "Python parent id": 1933, "Python id": 1934, "finished": true, "Ev Idx": 8764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564834886.225, "dur": 22.770,
    "args": {
      "Python parent id": 1933, "Python id": 1935, "finished": true, "Ev Idx": 8765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564834910.070, "dur": 32.732,
    "args": {
      "Python parent id": 1932, "Python id": 1936, "finished": true, "Ev Idx": 8766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564834924.714, "dur": 2.825,
    "args": {
      "Python parent id": 1936, "Python id": 1937, "finished": true, "Ev Idx": 8767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564834928.048, "dur": 14.680,
    "args": {
      "Python parent id": 1936, "Python id": 1938, "finished": true, "Ev Idx": 8768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564834945.205, "dur": 2.987,
    "args": {
      "Python parent id": 1932, "Python id": 1939, "finished": true, "Ev Idx": 8769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564834948.631, "dur": 50.612,
    "args": {
      "Python parent id": 1932, "Python id": 1940, "finished": true, "Ev Idx": 8770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564835032.644, "dur": 17.022,
    "args": {
      "Python parent id": 1932, "Python id": 1941, "finished": true, "Ev Idx": 8771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564835032.863, "dur": 0.324,
    "args": {
      "Python parent id": 1941, "Python id": 1942, "finished": true, "Ev Idx": 8772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564835033.995, "dur": 15.530,
    "args": {
      "Python parent id": 1941, "Python id": 1943, "finished": true, "Ev Idx": 8773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564835050.595, "dur": 2.047,
    "args": {
      "Python parent id": 1932, "Python id": 1944, "finished": true, "Ev Idx": 8774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564835054.206, "dur": 10.598,
    "args": {
      "Python parent id": 1932, "Python id": 1945, "finished": true, "Ev Idx": 8775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564835054.304, "dur": 0.099,
    "args": {
      "Python parent id": 1945, "Python id": 1946, "finished": true, "Ev Idx": 8776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835060.397, "dur": 1.066,
    "args": {
      "Python parent id": 1945, "Python id": 1947, "finished": true, "Ev Idx": 8777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564835060.934, "dur": 0.433,
    "args": {
      "Python parent id": 1947, "Python id": 1948, "finished": true, "Ev Idx": 8778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564835061.873, "dur": 2.861,
    "args": {
      "Python parent id": 1945, "Python id": 1949, "finished": true, "Ev Idx": 8779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564835065.573, "dur": 39.648,
    "args": {
      "Python parent id": 1932, "Python id": 1950, "finished": true, "Ev Idx": 8780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564835105.929, "dur": 3.417,
    "args": {
      "Python parent id": 1932, "Python id": 1951, "finished": true, "Ev Idx": 8781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564835110.099, "dur": 13.998,
    "args": {
      "Python parent id": 1932, "Python id": 1952, "finished": true, "Ev Idx": 8782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564835128.115, "dur": 3.440,
    "args": {
      "Python parent id": 1883, "Python id": 1953, "finished": true, "Ev Idx": 8783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564835132.249, "dur": 0.294,
    "args": {
      "Python parent id": 1883, "Python id": 1954, "finished": true, "Ev Idx": 8784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835133.614, "dur": 0.856,
    "args": {
      "Python parent id": 1883, "Python id": 1955, "finished": true, "Ev Idx": 8785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_73", "pid": 537, "tid": 537,
    "ts": 7429564835138.630, "dur": 44.210,
    "args": {
      "Python parent id": 1883, "Python module id": 73, "Python id": 1956, "finished": true, "Ev Idx": 8786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835139.665, "dur": 42.801,
    "args": {
      "Python parent id": 1956, "Python id": 1957, "finished": true, "Ev Idx": 8787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835140.053, "dur": 0.910,
    "args": {
      "Python parent id": 1957, "Python id": 1958, "finished": true, "Ev Idx": 8788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564835142.081, "dur": 40.124,
    "args": {
      "Python parent id": 1957, "Python id": 1959, "finished": true, "Ev Idx": 8789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835142.449, "dur": 0.158,
    "args": {
      "Python parent id": 1959, "Python id": 1960, "finished": true, "Ev Idx": 8790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835142.830, "dur": 0.144,
    "args": {
      "Python parent id": 1959, "Python id": 1961, "finished": true, "Ev Idx": 8791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564835143.165, "dur": 38.859,
    "args": {
      "Python parent id": 1959, "Python id": 1962, "finished": true, "Ev Idx": 8792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835197.346, "dur": 0.572,
    "args": {
      "Python parent id": 1867, "Python id": 1963, "finished": true, "Ev Idx": 8793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_21", "pid": 537, "tid": 537,
    "ts": 7429564835200.640, "dur": 81.488,
    "args": {
      "Python parent id": 1867, "Python module id": 21, "Python id": 1964, "finished": true, "Ev Idx": 8794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835201.255, "dur": 80.614,
    "args": {
      "Python parent id": 1964, "Python id": 1965, "finished": true, "Ev Idx": 8795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835201.478, "dur": 0.484,
    "args": {
      "Python parent id": 1965, "Python id": 1966, "finished": true, "Ev Idx": 8796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564835202.854, "dur": 78.241,
    "args": {
      "Python parent id": 1965, "Python id": 1967, "finished": true, "Ev Idx": 8797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564835203.383, "dur": 1.631,
    "args": {
      "Python parent id": 1967, "Python id": 1968, "finished": true, "Ev Idx": 8798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564835205.334, "dur": 13.901,
    "args": {
      "Python parent id": 1967, "Python id": 1969, "finished": true, "Ev Idx": 8799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564835219.597, "dur": 13.980,
    "args": {
      "Python parent id": 1967, "Python id": 1970, "finished": true, "Ev Idx": 8800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564835247.138, "dur": 12.136,
    "args": {
      "Python parent id": 1967, "Python id": 1971, "finished": true, "Ev Idx": 8801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835269.284, "dur": 0.311,
    "args": {
      "Python parent id": 1967, "Python id": 1972, "finished": true, "Ev Idx": 8802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564835269.965, "dur": 1.438,
    "args": {
      "Python parent id": 1967, "Python id": 1973, "finished": true, "Ev Idx": 8803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835282.920, "dur": 0.370,
    "args": {
      "Python parent id": 1867, "Python id": 1974, "finished": true, "Ev Idx": 8804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_10", "pid": 537, "tid": 537,
    "ts": 7429564835284.919, "dur": 283.361,
    "args": {
      "Python parent id": 1867, "Python module id": 10, "Python id": 1975, "finished": true, "Ev Idx": 8805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835285.457, "dur": 282.647,
    "args": {
      "Python parent id": 1975, "Python id": 1976, "finished": true, "Ev Idx": 8806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835285.673, "dur": 0.297,
    "args": {
      "Python parent id": 1976, "Python id": 1977, "finished": true, "Ev Idx": 8807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564835287.199, "dur": 280.799,
    "args": {
      "Python parent id": 1976, "Python id": 1978, "finished": true, "Ev Idx": 8808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835287.491, "dur": 0.421,
    "args": {
      "Python parent id": 1978, "Python id": 1979, "finished": true, "Ev Idx": 8809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835288.130, "dur": 0.296,
    "args": {
      "Python parent id": 1978, "Python id": 1980, "finished": true, "Ev Idx": 8810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835288.690, "dur": 0.265,
    "args": {
      "Python parent id": 1978, "Python id": 1981, "finished": true, "Ev Idx": 8811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_74", "pid": 537, "tid": 537,
    "ts": 7429564835290.934, "dur": 113.571,
    "args": {
      "Python parent id": 1978, "Python module id": 74, "Python id": 1982, "finished": true, "Ev Idx": 8812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835291.471, "dur": 112.697,
    "args": {
      "Python parent id": 1982, "Python id": 1983, "finished": true, "Ev Idx": 8813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835291.615, "dur": 0.201,
    "args": {
      "Python parent id": 1983, "Python id": 1984, "finished": true, "Ev Idx": 8814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564835292.563, "dur": 111.370,
    "args": {
      "Python parent id": 1983, "Python id": 1985, "finished": true, "Ev Idx": 8815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835292.851, "dur": 0.159,
    "args": {
      "Python parent id": 1985, "Python id": 1986, "finished": true, "Ev Idx": 8816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835293.259, "dur": 0.108,
    "args": {
      "Python parent id": 1985, "Python id": 1987, "finished": true, "Ev Idx": 8817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564835293.492, "dur": 110.249,
    "args": {
      "Python parent id": 1985, "Python id": 1988, "finished": true, "Ev Idx": 8818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_10", "pid": 537, "tid": 537,
    "ts": 7429564835406.671, "dur": 17.975,
    "args": {
      "Python parent id": 1978, "Python module id": 10, "Python id": 1989, "finished": true, "Ev Idx": 8819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835407.439, "dur": 16.992,
    "args": {
      "Python parent id": 1989, "Python id": 1990, "finished": true, "Ev Idx": 8820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835407.743, "dur": 0.396,
    "args": {
      "Python parent id": 1990, "Python id": 1991, "finished": true, "Ev Idx": 8821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564835409.196, "dur": 15.088,
    "args": {
      "Python parent id": 1990, "Python id": 1992, "finished": true, "Ev Idx": 8822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564835410.311, "dur": 13.818,
    "args": {
      "Python parent id": 1992, "Python id": 1993, "finished": true, "Ev Idx": 8823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564835410.475, "dur": 0.311,
    "args": {
      "Python parent id": 1993, "Python id": 1994, "finished": true, "Ev Idx": 8824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564835411.828, "dur": 12.203,
    "args": {
      "Python parent id": 1993, "Python id": 1995, "finished": true, "Ev Idx": 8825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835425.406, "dur": 0.522,
    "args": {
      "Python parent id": 1978, "Python id": 1996, "finished": true, "Ev Idx": 8826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_75", "pid": 537, "tid": 537,
    "ts": 7429564835428.149, "dur": 82.867,
    "args": {
      "Python parent id": 1978, "Python module id": 75, "Python id": 1997, "finished": true, "Ev Idx": 8827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835428.754, "dur": 82.033,
    "args": {
      "Python parent id": 1997, "Python id": 1998, "finished": true, "Ev Idx": 8828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835428.948, "dur": 0.219,
    "args": {
      "Python parent id": 1998, "Python id": 1999, "finished": true, "Ev Idx": 8829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564835429.837, "dur": 80.749,
    "args": {
      "Python parent id": 1998, "Python id": 2000, "finished": true, "Ev Idx": 8830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835430.130, "dur": 0.151,
    "args": {
      "Python parent id": 2000, "Python id": 2001, "finished": true, "Ev Idx": 8831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835430.480, "dur": 0.167,
    "args": {
      "Python parent id": 2000, "Python id": 2002, "finished": true, "Ev Idx": 8832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564835430.774, "dur": 79.705,
    "args": {
      "Python parent id": 2000, "Python id": 2003, "finished": true, "Ev Idx": 8833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_76", "pid": 537, "tid": 537,
    "ts": 7429564835524.536, "dur": 43.142,
    "args": {
      "Python parent id": 1978, "Python module id": 76, "Python id": 2004, "finished": true, "Ev Idx": 8834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835525.334, "dur": 42.055,
    "args": {
      "Python parent id": 2004, "Python id": 2005, "finished": true, "Ev Idx": 8835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835525.576, "dur": 0.405,
    "args": {
      "Python parent id": 2005, "Python id": 2006, "finished": true, "Ev Idx": 8836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564835526.742, "dur": 40.510,
    "args": {
      "Python parent id": 2005, "Python id": 2007, "finished": true, "Ev Idx": 8837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835527.062, "dur": 0.215,
    "args": {
      "Python parent id": 2007, "Python id": 2008, "finished": true, "Ev Idx": 8838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835527.492, "dur": 0.130,
    "args": {
      "Python parent id": 2007, "Python id": 2009, "finished": true, "Ev Idx": 8839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564835527.776, "dur": 39.371,
    "args": {
      "Python parent id": 2007, "Python id": 2010, "finished": true, "Ev Idx": 8840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_11", "pid": 537, "tid": 537,
    "ts": 7429564835588.678, "dur": 1222.200,
    "args": {
      "Python parent id": 64, "Python module id": 11, "Python id": 2011, "finished": true, "Ev Idx": 8841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835590.037, "dur": 1220.065,
    "args": {
      "Python parent id": 2011, "Python id": 2012, "finished": true, "Ev Idx": 8842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835590.242, "dur": 0.335,
    "args": {
      "Python parent id": 2012, "Python id": 2013, "finished": true, "Ev Idx": 8843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564835592.196, "dur": 1217.015,
    "args": {
      "Python parent id": 2012, "Python id": 2014, "finished": true, "Ev Idx": 8844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835592.663, "dur": 0.315,
    "args": {
      "Python parent id": 2014, "Python id": 2015, "finished": true, "Ev Idx": 8845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_22", "pid": 537, "tid": 537,
    "ts": 7429564835595.179, "dur": 78.133,
    "args": {
      "Python parent id": 2014, "Python module id": 22, "Python id": 2016, "finished": true, "Ev Idx": 8846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835595.700, "dur": 77.294,
    "args": {
      "Python parent id": 2016, "Python id": 2017, "finished": true, "Ev Idx": 8847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835595.839, "dur": 0.173,
    "args": {
      "Python parent id": 2017, "Python id": 2018, "finished": true, "Ev Idx": 8848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564835596.732, "dur": 75.602,
    "args": {
      "Python parent id": 2017, "Python id": 2019, "finished": true, "Ev Idx": 8849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564835597.211, "dur": 1.846,
    "args": {
      "Python parent id": 2019, "Python id": 2020, "finished": true, "Ev Idx": 8850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564835599.324, "dur": 12.807,
    "args": {
      "Python parent id": 2019, "Python id": 2021, "finished": true, "Ev Idx": 8851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564835612.524, "dur": 14.093,
    "args": {
      "Python parent id": 2019, "Python id": 2022, "finished": true, "Ev Idx": 8852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564835638.377, "dur": 11.201,
    "args": {
      "Python parent id": 2019, "Python id": 2023, "finished": true, "Ev Idx": 8853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835659.938, "dur": 0.187,
    "args": {
      "Python parent id": 2019, "Python id": 2024, "finished": true, "Ev Idx": 8854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564835660.681, "dur": 1.334,
    "args": {
      "Python parent id": 2019, "Python id": 2025, "finished": true, "Ev Idx": 8855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835673.967, "dur": 0.346,
    "args": {
      "Python parent id": 2014, "Python id": 2026, "finished": true, "Ev Idx": 8856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_11", "pid": 537, "tid": 537,
    "ts": 7429564835676.830, "dur": 732.087,
    "args": {
      "Python parent id": 2014, "Python module id": 11, "Python id": 2027, "finished": true, "Ev Idx": 8857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835678.018, "dur": 730.230,
    "args": {
      "Python parent id": 2027, "Python id": 2028, "finished": true, "Ev Idx": 8858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835678.194, "dur": 0.331,
    "args": {
      "Python parent id": 2028, "Python id": 2029, "finished": true, "Ev Idx": 8859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564835680.109, "dur": 727.200,
    "args": {
      "Python parent id": 2028, "Python id": 2030, "finished": true, "Ev Idx": 8860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835683.284, "dur": 0.326,
    "args": {
      "Python parent id": 2030, "Python id": 2031, "finished": true, "Ev Idx": 8861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_77", "pid": 537, "tid": 537,
    "ts": 7429564835685.713, "dur": 42.467,
    "args": {
      "Python parent id": 2030, "Python module id": 77, "Python id": 2032, "finished": true, "Ev Idx": 8862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835686.315, "dur": 41.593,
    "args": {
      "Python parent id": 2032, "Python id": 2033, "finished": true, "Ev Idx": 8863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835686.456, "dur": 0.180,
    "args": {
      "Python parent id": 2033, "Python id": 2034, "finished": true, "Ev Idx": 8864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564835687.434, "dur": 40.308,
    "args": {
      "Python parent id": 2033, "Python id": 2035, "finished": true, "Ev Idx": 8865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835687.654, "dur": 0.120,
    "args": {
      "Python parent id": 2035, "Python id": 2036, "finished": true, "Ev Idx": 8866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835687.988, "dur": 0.114,
    "args": {
      "Python parent id": 2035, "Python id": 2037, "finished": true, "Ev Idx": 8867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564835688.233, "dur": 39.439,
    "args": {
      "Python parent id": 2035, "Python id": 2038, "finished": true, "Ev Idx": 8868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564835728.754, "dur": 3.300,
    "args": {
      "Python parent id": 2030, "Python id": 2039, "finished": true, "Ev Idx": 8869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564835732.553, "dur": 3.521,
    "args": {
      "Python parent id": 2030, "Python id": 2040, "finished": true, "Ev Idx": 8870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835736.881, "dur": 0.372,
    "args": {
      "Python parent id": 2030, "Python id": 2041, "finished": true, "Ev Idx": 8871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_78", "pid": 537, "tid": 537,
    "ts": 7429564835739.294, "dur": 43.132,
    "args": {
      "Python parent id": 2030, "Python module id": 78, "Python id": 2042, "finished": true, "Ev Idx": 8872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835739.821, "dur": 42.380,
    "args": {
      "Python parent id": 2042, "Python id": 2043, "finished": true, "Ev Idx": 8873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835740.063, "dur": 0.294,
    "args": {
      "Python parent id": 2043, "Python id": 2044, "finished": true, "Ev Idx": 8874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564835741.121, "dur": 40.970,
    "args": {
      "Python parent id": 2043, "Python id": 2045, "finished": true, "Ev Idx": 8875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835741.335, "dur": 0.157,
    "args": {
      "Python parent id": 2045, "Python id": 2046, "finished": true, "Ev Idx": 8876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835741.687, "dur": 0.123,
    "args": {
      "Python parent id": 2045, "Python id": 2047, "finished": true, "Ev Idx": 8877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564835741.915, "dur": 40.120,
    "args": {
      "Python parent id": 2045, "Python id": 2048, "finished": true, "Ev Idx": 8878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564835782.912, "dur": 2.530,
    "args": {
      "Python parent id": 2030, "Python id": 2049, "finished": true, "Ev Idx": 8879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564835785.924, "dur": 2.697,
    "args": {
      "Python parent id": 2030, "Python id": 2050, "finished": true, "Ev Idx": 8880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835789.261, "dur": 0.429,
    "args": {
      "Python parent id": 2030, "Python id": 2051, "finished": true, "Ev Idx": 8881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_79", "pid": 537, "tid": 537,
    "ts": 7429564835791.845, "dur": 40.483,
    "args": {
      "Python parent id": 2030, "Python module id": 79, "Python id": 2052, "finished": true, "Ev Idx": 8882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564835792.436, "dur": 39.647,
    "args": {
      "Python parent id": 2052, "Python id": 2053, "finished": true, "Ev Idx": 8883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564835792.653, "dur": 0.303,
    "args": {
      "Python parent id": 2053, "Python id": 2054, "finished": true, "Ev Idx": 8884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564835793.631, "dur": 38.283,
    "args": {
      "Python parent id": 2053, "Python id": 2055, "finished": true, "Ev Idx": 8885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835793.901, "dur": 0.155,
    "args": {
      "Python parent id": 2055, "Python id": 2056, "finished": true, "Ev Idx": 8886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564835794.259, "dur": 0.150,
    "args": {
      "Python parent id": 2055, "Python id": 2057, "finished": true, "Ev Idx": 8887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564835794.536, "dur": 37.332,
    "args": {
      "Python parent id": 2055, "Python id": 2058, "finished": true, "Ev Idx": 8888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564835832.856, "dur": 2.488,
    "args": {
      "Python parent id": 2030, "Python id": 2059, "finished": true, "Ev Idx": 8889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564835836.035, "dur": 4.189,
    "args": {
      "Python parent id": 2030, "Python id": 2060, "finished": true, "Ev Idx": 8890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564835841.438, "dur": 190.982,
    "args": {
      "Python parent id": 2030, "Python id": 2061, "finished": true, "Ev Idx": 8891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564835841.736, "dur": 3.699,
    "args": {
      "Python parent id": 2061, "Python id": 2062, "finished": true, "Ev Idx": 8892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564835845.784, "dur": 1.803,
    "args": {
      "Python parent id": 2061, "Python id": 2063, "finished": true, "Ev Idx": 8893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564835860.285, "dur": 89.912,
    "args": {
      "Python parent id": 2061, "Python id": 2064, "finished": true, "Ev Idx": 8894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564835887.816, "dur": 61.164,
    "args": {
      "Python parent id": 2064, "Python id": 2065, "finished": true, "Ev Idx": 8895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564835981.741, "dur": 32.862,
    "args": {
      "Python parent id": 2061, "Python id": 2066, "finished": true, "Ev Idx": 8896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564835999.123, "dur": 14.621,
    "args": {
      "Python parent id": 2066, "Python id": 2067, "finished": true, "Ev Idx": 8897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564836036.992, "dur": 3.817,
    "args": {
      "Python parent id": 2030, "Python id": 2068, "finished": true, "Ev Idx": 8898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564836037.764, "dur": 0.359,
    "args": {
      "Python parent id": 2068, "Python id": 2069, "finished": true, "Ev Idx": 8899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564836038.428, "dur": 0.109,
    "args": {
      "Python parent id": 2068, "Python id": 2070, "finished": true, "Ev Idx": 8900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564836039.773, "dur": 0.231,
    "args": {
      "Python parent id": 2068, "Python id": 2071, "finished": true, "Ev Idx": 8901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564836040.389, "dur": 0.142,
    "args": {
      "Python parent id": 2068, "Python id": 2072, "finished": true, "Ev Idx": 8902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564836042.103, "dur": 6.595,
    "args": {
      "Python parent id": 2030, "Python id": 2073, "finished": true, "Ev Idx": 8903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564836045.251, "dur": 3.321,
    "args": {
      "Python parent id": 2073, "Python id": 2074, "finished": true, "Ev Idx": 8904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564836045.451, "dur": 1.460,
    "args": {
      "Python parent id": 2074, "Python id": 2075, "finished": true, "Ev Idx": 8905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564836045.978, "dur": 0.730,
    "args": {
      "Python parent id": 2075, "Python id": 2076, "finished": true, "Ev Idx": 8906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564836047.198, "dur": 0.545,
    "args": {
      "Python parent id": 2074, "Python id": 2077, "finished": true, "Ev Idx": 8907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564836048.068, "dur": 0.429,
    "args": {
      "Python parent id": 2074, "Python id": 2078, "finished": true, "Ev Idx": 8908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564836051.378, "dur": 295.618,
    "args": {
      "Python parent id": 2030, "Python id": 2079, "finished": true, "Ev Idx": 8909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564836051.919, "dur": 46.838,
    "args": {
      "Python parent id": 2079, "Python id": 2080, "finished": true, "Ev Idx": 8910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564836069.725, "dur": 5.582,
    "args": {
      "Python parent id": 2080, "Python id": 2081, "finished": true, "Ev Idx": 8911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564836076.226, "dur": 22.370,
    "args": {
      "Python parent id": 2080, "Python id": 2082, "finished": true, "Ev Idx": 8912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564836099.662, "dur": 31.872,
    "args": {
      "Python parent id": 2079, "Python id": 2083, "finished": true, "Ev Idx": 8913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564836115.617, "dur": 2.939,
    "args": {
      "Python parent id": 2083, "Python id": 2084, "finished": true, "Ev Idx": 8914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564836119.069, "dur": 12.402,
    "args": {
      "Python parent id": 2083, "Python id": 2085, "finished": true, "Ev Idx": 8915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564836132.394, "dur": 3.927,
    "args": {
      "Python parent id": 2079, "Python id": 2086, "finished": true, "Ev Idx": 8916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564836136.643, "dur": 49.508,
    "args": {
      "Python parent id": 2079, "Python id": 2087, "finished": true, "Ev Idx": 8917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564836221.105, "dur": 17.116,
    "args": {
      "Python parent id": 2079, "Python id": 2088, "finished": true, "Ev Idx": 8918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564836221.259, "dur": 0.334,
    "args": {
      "Python parent id": 2088, "Python id": 2089, "finished": true, "Ev Idx": 8919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564836222.446, "dur": 15.638,
    "args": {
      "Python parent id": 2088, "Python id": 2090, "finished": true, "Ev Idx": 8920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564836239.157, "dur": 1.930,
    "args": {
      "Python parent id": 2079, "Python id": 2091, "finished": true, "Ev Idx": 8921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564836242.655, "dur": 9.683,
    "args": {
      "Python parent id": 2079, "Python id": 2092, "finished": true, "Ev Idx": 8922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564836242.769, "dur": 0.104,
    "args": {
      "Python parent id": 2092, "Python id": 2093, "finished": true, "Ev Idx": 8923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836248.128, "dur": 0.940,
    "args": {
      "Python parent id": 2092, "Python id": 2094, "finished": true, "Ev Idx": 8924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564836248.548, "dur": 0.437,
    "args": {
      "Python parent id": 2094, "Python id": 2095, "finished": true, "Ev Idx": 8925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564836249.488, "dur": 2.765,
    "args": {
      "Python parent id": 2092, "Python id": 2096, "finished": true, "Ev Idx": 8926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564836253.059, "dur": 74.948,
    "args": {
      "Python parent id": 2079, "Python id": 2097, "finished": true, "Ev Idx": 8927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564836328.671, "dur": 3.119,
    "args": {
      "Python parent id": 2079, "Python id": 2098, "finished": true, "Ev Idx": 8928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564836332.510, "dur": 13.399,
    "args": {
      "Python parent id": 2079, "Python id": 2099, "finished": true, "Ev Idx": 8929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564836350.025, "dur": 3.450,
    "args": {
      "Python parent id": 2030, "Python id": 2100, "finished": true, "Ev Idx": 8930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564836354.188, "dur": 0.290,
    "args": {
      "Python parent id": 2030, "Python id": 2101, "finished": true, "Ev Idx": 8931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836355.525, "dur": 0.950,
    "args": {
      "Python parent id": 2030, "Python id": 2102, "finished": true, "Ev Idx": 8932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_80", "pid": 537, "tid": 537,
    "ts": 7429564836361.308, "dur": 45.397,
    "args": {
      "Python parent id": 2030, "Python module id": 80, "Python id": 2103, "finished": true, "Ev Idx": 8933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836362.578, "dur": 43.659,
    "args": {
      "Python parent id": 2103, "Python id": 2104, "finished": true, "Ev Idx": 8934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836362.960, "dur": 1.401,
    "args": {
      "Python parent id": 2104, "Python id": 2105, "finished": true, "Ev Idx": 8935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564836365.786, "dur": 40.129,
    "args": {
      "Python parent id": 2104, "Python id": 2106, "finished": true, "Ev Idx": 8936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836366.204, "dur": 0.250,
    "args": {
      "Python parent id": 2106, "Python id": 2107, "finished": true, "Ev Idx": 8937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836366.747, "dur": 0.171,
    "args": {
      "Python parent id": 2106, "Python id": 2108, "finished": true, "Ev Idx": 8938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564836367.165, "dur": 38.646,
    "args": {
      "Python parent id": 2106, "Python id": 2109, "finished": true, "Ev Idx": 8939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836422.004, "dur": 0.822,
    "args": {
      "Python parent id": 2014, "Python id": 2110, "finished": true, "Ev Idx": 8940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_23", "pid": 537, "tid": 537,
    "ts": 7429564836425.923, "dur": 86.115,
    "args": {
      "Python parent id": 2014, "Python module id": 23, "Python id": 2111, "finished": true, "Ev Idx": 8941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836426.637, "dur": 85.044,
    "args": {
      "Python parent id": 2111, "Python id": 2112, "finished": true, "Ev Idx": 8942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836426.840, "dur": 0.467,
    "args": {
      "Python parent id": 2112, "Python id": 2113, "finished": true, "Ev Idx": 8943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564836428.275, "dur": 82.556,
    "args": {
      "Python parent id": 2112, "Python id": 2114, "finished": true, "Ev Idx": 8944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564836428.788, "dur": 1.656,
    "args": {
      "Python parent id": 2114, "Python id": 2115, "finished": true, "Ev Idx": 8945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564836430.784, "dur": 14.647,
    "args": {
      "Python parent id": 2114, "Python id": 2116, "finished": true, "Ev Idx": 8946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564836445.831, "dur": 14.911,
    "args": {
      "Python parent id": 2114, "Python id": 2117, "finished": true, "Ev Idx": 8947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564836473.677, "dur": 12.700,
    "args": {
      "Python parent id": 2114, "Python id": 2118, "finished": true, "Ev Idx": 8948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836497.681, "dur": 0.296,
    "args": {
      "Python parent id": 2114, "Python id": 2119, "finished": true, "Ev Idx": 8949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564836498.359, "dur": 1.260,
    "args": {
      "Python parent id": 2114, "Python id": 2120, "finished": true, "Ev Idx": 8950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836512.797, "dur": 0.413,
    "args": {
      "Python parent id": 2014, "Python id": 2121, "finished": true, "Ev Idx": 8951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_11", "pid": 537, "tid": 537,
    "ts": 7429564836515.267, "dur": 281.660,
    "args": {
      "Python parent id": 2014, "Python module id": 11, "Python id": 2122, "finished": true, "Ev Idx": 8952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836515.807, "dur": 280.865,
    "args": {
      "Python parent id": 2122, "Python id": 2123, "finished": true, "Ev Idx": 8953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836515.993, "dur": 0.352,
    "args": {
      "Python parent id": 2123, "Python id": 2124, "finished": true, "Ev Idx": 8954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564836517.506, "dur": 279.064,
    "args": {
      "Python parent id": 2123, "Python id": 2125, "finished": true, "Ev Idx": 8955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836517.798, "dur": 0.318,
    "args": {
      "Python parent id": 2125, "Python id": 2126, "finished": true, "Ev Idx": 8956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836518.457, "dur": 0.312,
    "args": {
      "Python parent id": 2125, "Python id": 2127, "finished": true, "Ev Idx": 8957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836519.049, "dur": 0.279,
    "args": {
      "Python parent id": 2125, "Python id": 2128, "finished": true, "Ev Idx": 8958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_81", "pid": 537, "tid": 537,
    "ts": 7429564836521.141, "dur": 51.021,
    "args": {
      "Python parent id": 2125, "Python module id": 81, "Python id": 2129, "finished": true, "Ev Idx": 8959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836521.869, "dur": 49.919,
    "args": {
      "Python parent id": 2129, "Python id": 2130, "finished": true, "Ev Idx": 8960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836522.019, "dur": 0.168,
    "args": {
      "Python parent id": 2130, "Python id": 2131, "finished": true, "Ev Idx": 8961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564836522.956, "dur": 48.681,
    "args": {
      "Python parent id": 2130, "Python id": 2132, "finished": true, "Ev Idx": 8962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836523.198, "dur": 0.137,
    "args": {
      "Python parent id": 2132, "Python id": 2133, "finished": true, "Ev Idx": 8963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836523.564, "dur": 0.151,
    "args": {
      "Python parent id": 2132, "Python id": 2134, "finished": true, "Ev Idx": 8964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564836523.868, "dur": 47.662,
    "args": {
      "Python parent id": 2132, "Python id": 2135, "finished": true, "Ev Idx": 8965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_11", "pid": 537, "tid": 537,
    "ts": 7429564836573.815, "dur": 72.708,
    "args": {
      "Python parent id": 2125, "Python module id": 11, "Python id": 2136, "finished": true, "Ev Idx": 8966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836574.385, "dur": 71.800,
    "args": {
      "Python parent id": 2136, "Python id": 2137, "finished": true, "Ev Idx": 8967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836574.655, "dur": 0.323,
    "args": {
      "Python parent id": 2137, "Python id": 2138, "finished": true, "Ev Idx": 8968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564836575.973, "dur": 70.063,
    "args": {
      "Python parent id": 2137, "Python id": 2139, "finished": true, "Ev Idx": 8969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564836576.847, "dur": 68.972,
    "args": {
      "Python parent id": 2139, "Python id": 2140, "finished": true, "Ev Idx": 8970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564836577.005, "dur": 0.186,
    "args": {
      "Python parent id": 2140, "Python id": 2141, "finished": true, "Ev Idx": 8971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564836578.122, "dur": 67.565,
    "args": {
      "Python parent id": 2140, "Python id": 2142, "finished": true, "Ev Idx": 8972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836647.308, "dur": 0.389,
    "args": {
      "Python parent id": 2125, "Python id": 2143, "finished": true, "Ev Idx": 8973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_82", "pid": 537, "tid": 537,
    "ts": 7429564836649.845, "dur": 51.364,
    "args": {
      "Python parent id": 2125, "Python module id": 82, "Python id": 2144, "finished": true, "Ev Idx": 8974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836650.301, "dur": 50.571,
    "args": {
      "Python parent id": 2144, "Python id": 2145, "finished": true, "Ev Idx": 8975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836650.474, "dur": 0.350,
    "args": {
      "Python parent id": 2145, "Python id": 2146, "finished": true, "Ev Idx": 8976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564836651.494, "dur": 49.282,
    "args": {
      "Python parent id": 2145, "Python id": 2147, "finished": true, "Ev Idx": 8977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836651.821, "dur": 0.173,
    "args": {
      "Python parent id": 2147, "Python id": 2148, "finished": true, "Ev Idx": 8978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836652.226, "dur": 0.117,
    "args": {
      "Python parent id": 2147, "Python id": 2149, "finished": true, "Ev Idx": 8979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564836652.477, "dur": 48.171,
    "args": {
      "Python parent id": 2147, "Python id": 2150, "finished": true, "Ev Idx": 8980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_83", "pid": 537, "tid": 537,
    "ts": 7429564836746.981, "dur": 49.248,
    "args": {
      "Python parent id": 2125, "Python module id": 83, "Python id": 2151, "finished": true, "Ev Idx": 8981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836747.780, "dur": 48.188,
    "args": {
      "Python parent id": 2151, "Python id": 2152, "finished": true, "Ev Idx": 8982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836748.006, "dur": 0.432,
    "args": {
      "Python parent id": 2152, "Python id": 2153, "finished": true, "Ev Idx": 8983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564836749.143, "dur": 46.687,
    "args": {
      "Python parent id": 2152, "Python id": 2154, "finished": true, "Ev Idx": 8984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836749.401, "dur": 0.199,
    "args": {
      "Python parent id": 2154, "Python id": 2155, "finished": true, "Ev Idx": 8985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836749.814, "dur": 0.139,
    "args": {
      "Python parent id": 2154, "Python id": 2156, "finished": true, "Ev Idx": 8986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564836750.063, "dur": 45.690,
    "args": {
      "Python parent id": 2154, "Python id": 2157, "finished": true, "Ev Idx": 8987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_12", "pid": 537, "tid": 537,
    "ts": 7429564836817.450, "dur": 1133.981,
    "args": {
      "Python parent id": 64, "Python module id": 12, "Python id": 2158, "finished": true, "Ev Idx": 8988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836818.977, "dur": 1131.753,
    "args": {
      "Python parent id": 2158, "Python id": 2159, "finished": true, "Ev Idx": 8989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836819.145, "dur": 0.458,
    "args": {
      "Python parent id": 2159, "Python id": 2160, "finished": true, "Ev Idx": 8990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564836821.387, "dur": 1128.484,
    "args": {
      "Python parent id": 2159, "Python id": 2161, "finished": true, "Ev Idx": 8991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836821.950, "dur": 0.533,
    "args": {
      "Python parent id": 2161, "Python id": 2162, "finished": true, "Ev Idx": 8992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_24", "pid": 537, "tid": 537,
    "ts": 7429564836824.953, "dur": 81.078,
    "args": {
      "Python parent id": 2161, "Python module id": 24, "Python id": 2163, "finished": true, "Ev Idx": 8993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836825.480, "dur": 80.256,
    "args": {
      "Python parent id": 2163, "Python id": 2164, "finished": true, "Ev Idx": 8994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836825.631, "dur": 0.169,
    "args": {
      "Python parent id": 2164, "Python id": 2165, "finished": true, "Ev Idx": 8995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564836826.672, "dur": 78.419,
    "args": {
      "Python parent id": 2164, "Python id": 2166, "finished": true, "Ev Idx": 8996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564836827.295, "dur": 1.827,
    "args": {
      "Python parent id": 2166, "Python id": 2167, "finished": true, "Ev Idx": 8997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564836829.399, "dur": 13.501,
    "args": {
      "Python parent id": 2166, "Python id": 2168, "finished": true, "Ev Idx": 8998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564836843.304, "dur": 14.832,
    "args": {
      "Python parent id": 2166, "Python id": 2169, "finished": true, "Ev Idx": 8999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564836871.295, "dur": 10.510,
    "args": {
      "Python parent id": 2166, "Python id": 2170, "finished": true, "Ev Idx": 9000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836892.980, "dur": 0.252,
    "args": {
      "Python parent id": 2166, "Python id": 2171, "finished": true, "Ev Idx": 9001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564836893.580, "dur": 1.207,
    "args": {
      "Python parent id": 2166, "Python id": 2172, "finished": true, "Ev Idx": 9002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836906.715, "dur": 0.301,
    "args": {
      "Python parent id": 2161, "Python id": 2173, "finished": true, "Ev Idx": 9003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_12", "pid": 537, "tid": 537,
    "ts": 7429564836909.856, "dur": 647.624,
    "args": {
      "Python parent id": 2161, "Python module id": 12, "Python id": 2174, "finished": true, "Ev Idx": 9004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836911.270, "dur": 645.698,
    "args": {
      "Python parent id": 2174, "Python id": 2175, "finished": true, "Ev Idx": 9005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836911.481, "dur": 0.343,
    "args": {
      "Python parent id": 2175, "Python id": 2176, "finished": true, "Ev Idx": 9006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564836913.476, "dur": 642.934,
    "args": {
      "Python parent id": 2175, "Python id": 2177, "finished": true, "Ev Idx": 9007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836916.607, "dur": 0.296,
    "args": {
      "Python parent id": 2177, "Python id": 2178, "finished": true, "Ev Idx": 9008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_84", "pid": 537, "tid": 537,
    "ts": 7429564836918.767, "dur": 41.632,
    "args": {
      "Python parent id": 2177, "Python module id": 84, "Python id": 2179, "finished": true, "Ev Idx": 9009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836919.343, "dur": 40.718,
    "args": {
      "Python parent id": 2179, "Python id": 2180, "finished": true, "Ev Idx": 9010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836919.495, "dur": 0.350,
    "args": {
      "Python parent id": 2180, "Python id": 2181, "finished": true, "Ev Idx": 9011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564836920.479, "dur": 39.444,
    "args": {
      "Python parent id": 2180, "Python id": 2182, "finished": true, "Ev Idx": 9012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836920.814, "dur": 0.175,
    "args": {
      "Python parent id": 2182, "Python id": 2183, "finished": true, "Ev Idx": 9013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836921.227, "dur": 0.146,
    "args": {
      "Python parent id": 2182, "Python id": 2184, "finished": true, "Ev Idx": 9014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564836921.512, "dur": 38.255,
    "args": {
      "Python parent id": 2182, "Python id": 2185, "finished": true, "Ev Idx": 9015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564836961.068, "dur": 3.816,
    "args": {
      "Python parent id": 2177, "Python id": 2186, "finished": true, "Ev Idx": 9016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564836965.614, "dur": 4.250,
    "args": {
      "Python parent id": 2177, "Python id": 2187, "finished": true, "Ev Idx": 9017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836970.719, "dur": 0.333,
    "args": {
      "Python parent id": 2177, "Python id": 2188, "finished": true, "Ev Idx": 9018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_85", "pid": 537, "tid": 537,
    "ts": 7429564836973.232, "dur": 45.842,
    "args": {
      "Python parent id": 2177, "Python module id": 85, "Python id": 2189, "finished": true, "Ev Idx": 9019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564836973.875, "dur": 44.921,
    "args": {
      "Python parent id": 2189, "Python id": 2190, "finished": true, "Ev Idx": 9020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564836974.122, "dur": 0.225,
    "args": {
      "Python parent id": 2190, "Python id": 2191, "finished": true, "Ev Idx": 9021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564836974.931, "dur": 43.741,
    "args": {
      "Python parent id": 2190, "Python id": 2192, "finished": true, "Ev Idx": 9022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836975.116, "dur": 0.298,
    "args": {
      "Python parent id": 2192, "Python id": 2193, "finished": true, "Ev Idx": 9023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564836975.607, "dur": 0.115,
    "args": {
      "Python parent id": 2192, "Python id": 2194, "finished": true, "Ev Idx": 9024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564836975.837, "dur": 42.780,
    "args": {
      "Python parent id": 2192, "Python id": 2195, "finished": true, "Ev Idx": 9025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564837019.590, "dur": 2.456,
    "args": {
      "Python parent id": 2177, "Python id": 2196, "finished": true, "Ev Idx": 9026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564837022.742, "dur": 2.793,
    "args": {
      "Python parent id": 2177, "Python id": 2197, "finished": true, "Ev Idx": 9027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837026.280, "dur": 0.398,
    "args": {
      "Python parent id": 2177, "Python id": 2198, "finished": true, "Ev Idx": 9028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_86", "pid": 537, "tid": 537,
    "ts": 7429564837028.513, "dur": 43.687,
    "args": {
      "Python parent id": 2177, "Python module id": 86, "Python id": 2199, "finished": true, "Ev Idx": 9029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564837028.918, "dur": 43.037,
    "args": {
      "Python parent id": 2199, "Python id": 2200, "finished": true, "Ev Idx": 9030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564837029.093, "dur": 0.371,
    "args": {
      "Python parent id": 2200, "Python id": 2201, "finished": true, "Ev Idx": 9031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564837030.056, "dur": 41.691,
    "args": {
      "Python parent id": 2200, "Python id": 2202, "finished": true, "Ev Idx": 9032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837030.312, "dur": 0.144,
    "args": {
      "Python parent id": 2202, "Python id": 2203, "finished": true, "Ev Idx": 9033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837030.650, "dur": 0.113,
    "args": {
      "Python parent id": 2202, "Python id": 2204, "finished": true, "Ev Idx": 9034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564837030.866, "dur": 40.826,
    "args": {
      "Python parent id": 2202, "Python id": 2205, "finished": true, "Ev Idx": 9035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564837072.760, "dur": 2.482,
    "args": {
      "Python parent id": 2177, "Python id": 2206, "finished": true, "Ev Idx": 9036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564837075.736, "dur": 2.647,
    "args": {
      "Python parent id": 2177, "Python id": 2207, "finished": true, "Ev Idx": 9037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564837079.719, "dur": 147.944,
    "args": {
      "Python parent id": 2177, "Python id": 2208, "finished": true, "Ev Idx": 9038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564837080.090, "dur": 3.551,
    "args": {
      "Python parent id": 2208, "Python id": 2209, "finished": true, "Ev Idx": 9039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564837083.914, "dur": 1.789,
    "args": {
      "Python parent id": 2208, "Python id": 2210, "finished": true, "Ev Idx": 9040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564837097.737, "dur": 50.642,
    "args": {
      "Python parent id": 2208, "Python id": 2211, "finished": true, "Ev Idx": 9041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564837125.684, "dur": 21.709,
    "args": {
      "Python parent id": 2211, "Python id": 2212, "finished": true, "Ev Idx": 9042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564837180.071, "dur": 27.691,
    "args": {
      "Python parent id": 2208, "Python id": 2213, "finished": true, "Ev Idx": 9043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564837196.270, "dur": 10.850,
    "args": {
      "Python parent id": 2213, "Python id": 2214, "finished": true, "Ev Idx": 9044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564837231.815, "dur": 3.936,
    "args": {
      "Python parent id": 2177, "Python id": 2215, "finished": true, "Ev Idx": 9045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564837232.538, "dur": 0.329,
    "args": {
      "Python parent id": 2215, "Python id": 2216, "finished": true, "Ev Idx": 9046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564837233.244, "dur": 0.062,
    "args": {
      "Python parent id": 2215, "Python id": 2217, "finished": true, "Ev Idx": 9047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564837234.786, "dur": 0.208,
    "args": {
      "Python parent id": 2215, "Python id": 2218, "finished": true, "Ev Idx": 9048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564837235.357, "dur": 0.128,
    "args": {
      "Python parent id": 2215, "Python id": 2219, "finished": true, "Ev Idx": 9049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564837237.185, "dur": 6.892,
    "args": {
      "Python parent id": 2177, "Python id": 2220, "finished": true, "Ev Idx": 9050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564837240.458, "dur": 3.478,
    "args": {
      "Python parent id": 2220, "Python id": 2221, "finished": true, "Ev Idx": 9051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564837240.635, "dur": 1.361,
    "args": {
      "Python parent id": 2221, "Python id": 2222, "finished": true, "Ev Idx": 9052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564837241.072, "dur": 0.714,
    "args": {
      "Python parent id": 2222, "Python id": 2223, "finished": true, "Ev Idx": 9053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564837242.287, "dur": 0.532,
    "args": {
      "Python parent id": 2221, "Python id": 2224, "finished": true, "Ev Idx": 9054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564837243.187, "dur": 0.670,
    "args": {
      "Python parent id": 2221, "Python id": 2225, "finished": true, "Ev Idx": 9055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564837246.800, "dur": 253.718,
    "args": {
      "Python parent id": 2177, "Python id": 2226, "finished": true, "Ev Idx": 9056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564837247.386, "dur": 46.173,
    "args": {
      "Python parent id": 2226, "Python id": 2227, "finished": true, "Ev Idx": 9057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564837266.300, "dur": 4.816,
    "args": {
      "Python parent id": 2227, "Python id": 2228, "finished": true, "Ev Idx": 9058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564837272.007, "dur": 21.430,
    "args": {
      "Python parent id": 2227, "Python id": 2229, "finished": true, "Ev Idx": 9059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564837294.487, "dur": 31.422,
    "args": {
      "Python parent id": 2226, "Python id": 2230, "finished": true, "Ev Idx": 9060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564837308.935, "dur": 2.998,
    "args": {
      "Python parent id": 2230, "Python id": 2231, "finished": true, "Ev Idx": 9061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564837312.456, "dur": 13.376,
    "args": {
      "Python parent id": 2230, "Python id": 2232, "finished": true, "Ev Idx": 9062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564837326.665, "dur": 2.828,
    "args": {
      "Python parent id": 2226, "Python id": 2233, "finished": true, "Ev Idx": 9063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564837329.777, "dur": 44.312,
    "args": {
      "Python parent id": 2226, "Python id": 2234, "finished": true, "Ev Idx": 9064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564837407.148, "dur": 16.455,
    "args": {
      "Python parent id": 2226, "Python id": 2235, "finished": true, "Ev Idx": 9065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564837407.297, "dur": 0.251,
    "args": {
      "Python parent id": 2235, "Python id": 2236, "finished": true, "Ev Idx": 9066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564837408.211, "dur": 15.233,
    "args": {
      "Python parent id": 2235, "Python id": 2237, "finished": true, "Ev Idx": 9067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564837424.353, "dur": 1.615,
    "args": {
      "Python parent id": 2226, "Python id": 2238, "finished": true, "Ev Idx": 9068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564837427.397, "dur": 10.713,
    "args": {
      "Python parent id": 2226, "Python id": 2239, "finished": true, "Ev Idx": 9069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564837427.529, "dur": 0.120,
    "args": {
      "Python parent id": 2239, "Python id": 2240, "finished": true, "Ev Idx": 9070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837434.105, "dur": 0.838,
    "args": {
      "Python parent id": 2239, "Python id": 2241, "finished": true, "Ev Idx": 9071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564837434.550, "dur": 0.291,
    "args": {
      "Python parent id": 2241, "Python id": 2242, "finished": true, "Ev Idx": 9072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564837435.349, "dur": 2.693,
    "args": {
      "Python parent id": 2239, "Python id": 2243, "finished": true, "Ev Idx": 9073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564837438.728, "dur": 43.133,
    "args": {
      "Python parent id": 2226, "Python id": 2244, "finished": true, "Ev Idx": 9074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564837482.453, "dur": 3.100,
    "args": {
      "Python parent id": 2226, "Python id": 2245, "finished": true, "Ev Idx": 9075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564837486.284, "dur": 13.229,
    "args": {
      "Python parent id": 2226, "Python id": 2246, "finished": true, "Ev Idx": 9076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564837503.179, "dur": 3.157,
    "args": {
      "Python parent id": 2177, "Python id": 2247, "finished": true, "Ev Idx": 9077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564837506.997, "dur": 0.309,
    "args": {
      "Python parent id": 2177, "Python id": 2248, "finished": true, "Ev Idx": 9078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837508.280, "dur": 0.853,
    "args": {
      "Python parent id": 2177, "Python id": 2249, "finished": true, "Ev Idx": 9079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_87", "pid": 537, "tid": 537,
    "ts": 7429564837513.269, "dur": 42.600,
    "args": {
      "Python parent id": 2177, "Python module id": 87, "Python id": 2250, "finished": true, "Ev Idx": 9080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564837514.311, "dur": 41.173,
    "args": {
      "Python parent id": 2250, "Python id": 2251, "finished": true, "Ev Idx": 9081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564837514.679, "dur": 0.884,
    "args": {
      "Python parent id": 2251, "Python id": 2252, "finished": true, "Ev Idx": 9082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564837516.607, "dur": 38.610,
    "args": {
      "Python parent id": 2251, "Python id": 2253, "finished": true, "Ev Idx": 9083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837517.024, "dur": 0.197,
    "args": {
      "Python parent id": 2253, "Python id": 2254, "finished": true, "Ev Idx": 9084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837517.495, "dur": 0.131,
    "args": {
      "Python parent id": 2253, "Python id": 2255, "finished": true, "Ev Idx": 9085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564837517.779, "dur": 37.358,
    "args": {
      "Python parent id": 2253, "Python id": 2256, "finished": true, "Ev Idx": 9086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837569.483, "dur": 0.675,
    "args": {
      "Python parent id": 2161, "Python id": 2257, "finished": true, "Ev Idx": 9087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_25", "pid": 537, "tid": 537,
    "ts": 7429564837572.785, "dur": 133.522,
    "args": {
      "Python parent id": 2161, "Python module id": 25, "Python id": 2258, "finished": true, "Ev Idx": 9088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564837573.479, "dur": 132.348,
    "args": {
      "Python parent id": 2258, "Python id": 2259, "finished": true, "Ev Idx": 9089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564837573.676, "dur": 0.387,
    "args": {
      "Python parent id": 2259, "Python id": 2260, "finished": true, "Ev Idx": 9090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564837574.851, "dur": 129.854,
    "args": {
      "Python parent id": 2259, "Python id": 2261, "finished": true, "Ev Idx": 9091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564837575.333, "dur": 1.529,
    "args": {
      "Python parent id": 2261, "Python id": 2262, "finished": true, "Ev Idx": 9092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564837577.163, "dur": 13.328,
    "args": {
      "Python parent id": 2261, "Python id": 2263, "finished": true, "Ev Idx": 9093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564837590.818, "dur": 15.182,
    "args": {
      "Python parent id": 2261, "Python id": 2264, "finished": true, "Ev Idx": 9094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564837618.537, "dur": 9.489,
    "args": {
      "Python parent id": 2261, "Python id": 2265, "finished": true, "Ev Idx": 9095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837638.897, "dur": 0.206,
    "args": {
      "Python parent id": 2261, "Python id": 2266, "finished": true, "Ev Idx": 9096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564837639.471, "dur": 1.308,
    "args": {
      "Python parent id": 2261, "Python id": 2267, "finished": true, "Ev Idx": 9097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837707.015, "dur": 0.417,
    "args": {
      "Python parent id": 2161, "Python id": 2268, "finished": true, "Ev Idx": 9098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_12", "pid": 537, "tid": 537,
    "ts": 7429564837709.106, "dur": 228.722,
    "args": {
      "Python parent id": 2161, "Python module id": 12, "Python id": 2269, "finished": true, "Ev Idx": 9099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564837709.914, "dur": 227.712,
    "args": {
      "Python parent id": 2269, "Python id": 2270, "finished": true, "Ev Idx": 9100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564837710.127, "dur": 0.538,
    "args": {
      "Python parent id": 2270, "Python id": 2271, "finished": true, "Ev Idx": 9101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564837711.902, "dur": 225.586,
    "args": {
      "Python parent id": 2270, "Python id": 2272, "finished": true, "Ev Idx": 9102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837712.180, "dur": 0.336,
    "args": {
      "Python parent id": 2272, "Python id": 2273, "finished": true, "Ev Idx": 9103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837712.834, "dur": 0.291,
    "args": {
      "Python parent id": 2272, "Python id": 2274, "finished": true, "Ev Idx": 9104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837713.331, "dur": 0.260,
    "args": {
      "Python parent id": 2272, "Python id": 2275, "finished": true, "Ev Idx": 9105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_88", "pid": 537, "tid": 537,
    "ts": 7429564837716.004, "dur": 55.954,
    "args": {
      "Python parent id": 2272, "Python module id": 88, "Python id": 2276, "finished": true, "Ev Idx": 9106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564837716.725, "dur": 55.015,
    "args": {
      "Python parent id": 2276, "Python id": 2277, "finished": true, "Ev Idx": 9107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564837716.979, "dur": 0.178,
    "args": {
      "Python parent id": 2277, "Python id": 2278, "finished": true, "Ev Idx": 9108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564837717.784, "dur": 53.843,
    "args": {
      "Python parent id": 2277, "Python id": 2279, "finished": true, "Ev Idx": 9109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837718.068, "dur": 0.130,
    "args": {
      "Python parent id": 2279, "Python id": 2280, "finished": true, "Ev Idx": 9110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837718.434, "dur": 0.316,
    "args": {
      "Python parent id": 2279, "Python id": 2281, "finished": true, "Ev Idx": 9111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564837718.872, "dur": 52.609,
    "args": {
      "Python parent id": 2279, "Python id": 2282, "finished": true, "Ev Idx": 9112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_12", "pid": 537, "tid": 537,
    "ts": 7429564837773.984, "dur": 16.811,
    "args": {
      "Python parent id": 2272, "Python module id": 12, "Python id": 2283, "finished": true, "Ev Idx": 9113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564837774.553, "dur": 16.048,
    "args": {
      "Python parent id": 2283, "Python id": 2284, "finished": true, "Ev Idx": 9114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564837774.751, "dur": 0.354,
    "args": {
      "Python parent id": 2284, "Python id": 2285, "finished": true, "Ev Idx": 9115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564837776.102, "dur": 14.419,
    "args": {
      "Python parent id": 2284, "Python id": 2286, "finished": true, "Ev Idx": 9116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564837776.861, "dur": 13.491,
    "args": {
      "Python parent id": 2286, "Python id": 2287, "finished": true, "Ev Idx": 9117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564837776.989, "dur": 0.236,
    "args": {
      "Python parent id": 2287, "Python id": 2288, "finished": true, "Ev Idx": 9118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564837778.135, "dur": 12.125,
    "args": {
      "Python parent id": 2287, "Python id": 2289, "finished": true, "Ev Idx": 9119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837791.347, "dur": 0.340,
    "args": {
      "Python parent id": 2272, "Python id": 2290, "finished": true, "Ev Idx": 9120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_89", "pid": 537, "tid": 537,
    "ts": 7429564837793.551, "dur": 89.659,
    "args": {
      "Python parent id": 2272, "Python module id": 89, "Python id": 2291, "finished": true, "Ev Idx": 9121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564837793.919, "dur": 89.055,
    "args": {
      "Python parent id": 2291, "Python id": 2292, "finished": true, "Ev Idx": 9122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564837794.114, "dur": 0.322,
    "args": {
      "Python parent id": 2292, "Python id": 2293, "finished": true, "Ev Idx": 9123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564837795.097, "dur": 87.696,
    "args": {
      "Python parent id": 2292, "Python id": 2294, "finished": true, "Ev Idx": 9124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837795.306, "dur": 0.128,
    "args": {
      "Python parent id": 2294, "Python id": 2295, "finished": true, "Ev Idx": 9125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837795.641, "dur": 0.128,
    "args": {
      "Python parent id": 2294, "Python id": 2296, "finished": true, "Ev Idx": 9126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564837795.868, "dur": 86.815,
    "args": {
      "Python parent id": 2294, "Python id": 2297, "finished": true, "Ev Idx": 9127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_90", "pid": 537, "tid": 537,
    "ts": 7429564837895.758, "dur": 41.398,
    "args": {
      "Python parent id": 2272, "Python module id": 90, "Python id": 2298, "finished": true, "Ev Idx": 9128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564837896.361, "dur": 40.554,
    "args": {
      "Python parent id": 2298, "Python id": 2299, "finished": true, "Ev Idx": 9129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564837896.626, "dur": 0.401,
    "args": {
      "Python parent id": 2299, "Python id": 2300, "finished": true, "Ev Idx": 9130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564837897.818, "dur": 38.979,
    "args": {
      "Python parent id": 2299, "Python id": 2301, "finished": true, "Ev Idx": 9131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837898.110, "dur": 0.217,
    "args": {
      "Python parent id": 2301, "Python id": 2302, "finished": true, "Ev Idx": 9132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837898.546, "dur": 0.117,
    "args": {
      "Python parent id": 2301, "Python id": 2303, "finished": true, "Ev Idx": 9133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564837898.793, "dur": 37.924,
    "args": {
      "Python parent id": 2301, "Python id": 2304, "finished": true, "Ev Idx": 9134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_13", "pid": 537, "tid": 537,
    "ts": 7429564837957.838, "dur": 1229.839,
    "args": {
      "Python parent id": 64, "Python module id": 13, "Python id": 2305, "finished": true, "Ev Idx": 9135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564837959.276, "dur": 1227.671,
    "args": {
      "Python parent id": 2305, "Python id": 2306, "finished": true, "Ev Idx": 9136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564837959.491, "dur": 0.385,
    "args": {
      "Python parent id": 2306, "Python id": 2307, "finished": true, "Ev Idx": 9137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564837961.584, "dur": 1224.612,
    "args": {
      "Python parent id": 2306, "Python id": 2308, "finished": true, "Ev Idx": 9138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564837962.121, "dur": 0.680,
    "args": {
      "Python parent id": 2308, "Python id": 2309, "finished": true, "Ev Idx": 9139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_26", "pid": 537, "tid": 537,
    "ts": 7429564837964.631, "dur": 74.579,
    "args": {
      "Python parent id": 2308, "Python module id": 26, "Python id": 2310, "finished": true, "Ev Idx": 9140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564837965.039, "dur": 73.927,
    "args": {
      "Python parent id": 2310, "Python id": 2311, "finished": true, "Ev Idx": 9141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564837965.188, "dur": 0.154,
    "args": {
      "Python parent id": 2311, "Python id": 2312, "finished": true, "Ev Idx": 9142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564837966.162, "dur": 72.163,
    "args": {
      "Python parent id": 2311, "Python id": 2313, "finished": true, "Ev Idx": 9143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564837966.851, "dur": 1.916,
    "args": {
      "Python parent id": 2313, "Python id": 2314, "finished": true, "Ev Idx": 9144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564837969.106, "dur": 11.997,
    "args": {
      "Python parent id": 2313, "Python id": 2315, "finished": true, "Ev Idx": 9145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564837981.457, "dur": 13.957,
    "args": {
      "Python parent id": 2313, "Python id": 2316, "finished": true, "Ev Idx": 9146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564838007.158, "dur": 9.489,
    "args": {
      "Python parent id": 2313, "Python id": 2317, "finished": true, "Ev Idx": 9147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838026.870, "dur": 0.257,
    "args": {
      "Python parent id": 2313, "Python id": 2318, "finished": true, "Ev Idx": 9148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564838027.527, "dur": 1.167,
    "args": {
      "Python parent id": 2313, "Python id": 2319, "finished": true, "Ev Idx": 9149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838039.897, "dur": 0.346,
    "args": {
      "Python parent id": 2308, "Python id": 2320, "finished": true, "Ev Idx": 9150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_13", "pid": 537, "tid": 537,
    "ts": 7429564838045.106, "dur": 748.651,
    "args": {
      "Python parent id": 2308, "Python module id": 13, "Python id": 2321, "finished": true, "Ev Idx": 9151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564838046.602, "dur": 746.444,
    "args": {
      "Python parent id": 2321, "Python id": 2322, "finished": true, "Ev Idx": 9152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564838046.796, "dur": 0.300,
    "args": {
      "Python parent id": 2322, "Python id": 2323, "finished": true, "Ev Idx": 9153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564838048.853, "dur": 743.502,
    "args": {
      "Python parent id": 2322, "Python id": 2324, "finished": true, "Ev Idx": 9154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838051.793, "dur": 0.582,
    "args": {
      "Python parent id": 2324, "Python id": 2325, "finished": true, "Ev Idx": 9155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_91", "pid": 537, "tid": 537,
    "ts": 7429564838054.390, "dur": 109.588,
    "args": {
      "Python parent id": 2324, "Python module id": 91, "Python id": 2326, "finished": true, "Ev Idx": 9156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564838054.977, "dur": 108.652,
    "args": {
      "Python parent id": 2326, "Python id": 2327, "finished": true, "Ev Idx": 9157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564838055.122, "dur": 0.155,
    "args": {
      "Python parent id": 2327, "Python id": 2328, "finished": true, "Ev Idx": 9158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564838056.100, "dur": 107.224,
    "args": {
      "Python parent id": 2327, "Python id": 2329, "finished": true, "Ev Idx": 9159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838056.488, "dur": 0.135,
    "args": {
      "Python parent id": 2329, "Python id": 2330, "finished": true, "Ev Idx": 9160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838056.834, "dur": 0.130,
    "args": {
      "Python parent id": 2329, "Python id": 2331, "finished": true, "Ev Idx": 9161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564838057.086, "dur": 106.074,
    "args": {
      "Python parent id": 2329, "Python id": 2332, "finished": true, "Ev Idx": 9162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564838165.021, "dur": 4.787,
    "args": {
      "Python parent id": 2324, "Python id": 2333, "finished": true, "Ev Idx": 9163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564838170.366, "dur": 5.350,
    "args": {
      "Python parent id": 2324, "Python id": 2334, "finished": true, "Ev Idx": 9164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838177.045, "dur": 0.574,
    "args": {
      "Python parent id": 2324, "Python id": 2335, "finished": true, "Ev Idx": 9165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_92", "pid": 537, "tid": 537,
    "ts": 7429564838180.445, "dur": 50.512,
    "args": {
      "Python parent id": 2324, "Python module id": 92, "Python id": 2336, "finished": true, "Ev Idx": 9166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564838181.128, "dur": 49.540,
    "args": {
      "Python parent id": 2336, "Python id": 2337, "finished": true, "Ev Idx": 9167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564838181.442, "dur": 0.493,
    "args": {
      "Python parent id": 2337, "Python id": 2338, "finished": true, "Ev Idx": 9168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564838182.634, "dur": 47.948,
    "args": {
      "Python parent id": 2337, "Python id": 2339, "finished": true, "Ev Idx": 9169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838182.934, "dur": 0.141,
    "args": {
      "Python parent id": 2339, "Python id": 2340, "finished": true, "Ev Idx": 9170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838183.299, "dur": 0.128,
    "args": {
      "Python parent id": 2339, "Python id": 2341, "finished": true, "Ev Idx": 9171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564838183.562, "dur": 46.924,
    "args": {
      "Python parent id": 2339, "Python id": 2342, "finished": true, "Ev Idx": 9172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564838231.525, "dur": 2.711,
    "args": {
      "Python parent id": 2324, "Python id": 2343, "finished": true, "Ev Idx": 9173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564838234.724, "dur": 2.908,
    "args": {
      "Python parent id": 2324, "Python id": 2344, "finished": true, "Ev Idx": 9174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838238.760, "dur": 0.648,
    "args": {
      "Python parent id": 2324, "Python id": 2345, "finished": true, "Ev Idx": 9175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_93", "pid": 537, "tid": 537,
    "ts": 7429564838241.566, "dur": 44.084,
    "args": {
      "Python parent id": 2324, "Python module id": 93, "Python id": 2346, "finished": true, "Ev Idx": 9176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564838242.322, "dur": 43.077,
    "args": {
      "Python parent id": 2346, "Python id": 2347, "finished": true, "Ev Idx": 9177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564838242.534, "dur": 0.276,
    "args": {
      "Python parent id": 2347, "Python id": 2348, "finished": true, "Ev Idx": 9178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564838243.545, "dur": 41.717,
    "args": {
      "Python parent id": 2347, "Python id": 2349, "finished": true, "Ev Idx": 9179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838244.038, "dur": 0.138,
    "args": {
      "Python parent id": 2349, "Python id": 2350, "finished": true, "Ev Idx": 9180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838244.385, "dur": 0.130,
    "args": {
      "Python parent id": 2349, "Python id": 2351, "finished": true, "Ev Idx": 9181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564838244.618, "dur": 40.596,
    "args": {
      "Python parent id": 2349, "Python id": 2352, "finished": true, "Ev Idx": 9182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564838286.347, "dur": 2.476,
    "args": {
      "Python parent id": 2324, "Python id": 2353, "finished": true, "Ev Idx": 9183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564838289.305, "dur": 2.464,
    "args": {
      "Python parent id": 2324, "Python id": 2354, "finished": true, "Ev Idx": 9184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564838293.010, "dur": 159.345,
    "args": {
      "Python parent id": 2324, "Python id": 2355, "finished": true, "Ev Idx": 9185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564838293.344, "dur": 4.079,
    "args": {
      "Python parent id": 2355, "Python id": 2356, "finished": true, "Ev Idx": 9186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564838297.764, "dur": 1.998,
    "args": {
      "Python parent id": 2355, "Python id": 2357, "finished": true, "Ev Idx": 9187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564838315.411, "dur": 54.934,
    "args": {
      "Python parent id": 2355, "Python id": 2358, "finished": true, "Ev Idx": 9188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564838345.837, "dur": 23.372,
    "args": {
      "Python parent id": 2358, "Python id": 2359, "finished": true, "Ev Idx": 9189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564838402.437, "dur": 30.884,
    "args": {
      "Python parent id": 2355, "Python id": 2360, "finished": true, "Ev Idx": 9190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564838420.837, "dur": 11.506,
    "args": {
      "Python parent id": 2360, "Python id": 2361, "finished": true, "Ev Idx": 9191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564838457.050, "dur": 4.092,
    "args": {
      "Python parent id": 2324, "Python id": 2362, "finished": true, "Ev Idx": 9192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564838457.845, "dur": 0.340,
    "args": {
      "Python parent id": 2362, "Python id": 2363, "finished": true, "Ev Idx": 9193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564838458.562, "dur": 0.099,
    "args": {
      "Python parent id": 2362, "Python id": 2364, "finished": true, "Ev Idx": 9194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564838460.163, "dur": 0.224,
    "args": {
      "Python parent id": 2362, "Python id": 2365, "finished": true, "Ev Idx": 9195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564838460.788, "dur": 0.113,
    "args": {
      "Python parent id": 2362, "Python id": 2366, "finished": true, "Ev Idx": 9196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564838462.661, "dur": 6.620,
    "args": {
      "Python parent id": 2324, "Python id": 2367, "finished": true, "Ev Idx": 9197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564838465.872, "dur": 3.288,
    "args": {
      "Python parent id": 2367, "Python id": 2368, "finished": true, "Ev Idx": 9198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564838466.119, "dur": 1.405,
    "args": {
      "Python parent id": 2368, "Python id": 2369, "finished": true, "Ev Idx": 9199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564838466.587, "dur": 0.762,
    "args": {
      "Python parent id": 2369, "Python id": 2370, "finished": true, "Ev Idx": 9200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564838467.821, "dur": 0.487,
    "args": {
      "Python parent id": 2368, "Python id": 2371, "finished": true, "Ev Idx": 9201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564838468.600, "dur": 0.475,
    "args": {
      "Python parent id": 2368, "Python id": 2372, "finished": true, "Ev Idx": 9202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564838472.186, "dur": 263.592,
    "args": {
      "Python parent id": 2324, "Python id": 2373, "finished": true, "Ev Idx": 9203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564838472.816, "dur": 49.222,
    "args": {
      "Python parent id": 2373, "Python id": 2374, "finished": true, "Ev Idx": 9204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564838493.139, "dur": 5.336,
    "args": {
      "Python parent id": 2374, "Python id": 2375, "finished": true, "Ev Idx": 9205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564838499.422, "dur": 22.483,
    "args": {
      "Python parent id": 2374, "Python id": 2376, "finished": true, "Ev Idx": 9206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564838522.957, "dur": 32.454,
    "args": {
      "Python parent id": 2373, "Python id": 2377, "finished": true, "Ev Idx": 9207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564838537.294, "dur": 2.821,
    "args": {
      "Python parent id": 2377, "Python id": 2378, "finished": true, "Ev Idx": 9208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564838540.582, "dur": 14.766,
    "args": {
      "Python parent id": 2377, "Python id": 2379, "finished": true, "Ev Idx": 9209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564838556.315, "dur": 2.857,
    "args": {
      "Python parent id": 2373, "Python id": 2380, "finished": true, "Ev Idx": 9210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564838559.418, "dur": 48.780,
    "args": {
      "Python parent id": 2373, "Python id": 2381, "finished": true, "Ev Idx": 9211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564838641.165, "dur": 17.513,
    "args": {
      "Python parent id": 2373, "Python id": 2382, "finished": true, "Ev Idx": 9212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564838641.366, "dur": 0.289,
    "args": {
      "Python parent id": 2382, "Python id": 2383, "finished": true, "Ev Idx": 9213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564838642.347, "dur": 16.164,
    "args": {
      "Python parent id": 2382, "Python id": 2384, "finished": true, "Ev Idx": 9214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564838659.614, "dur": 1.909,
    "args": {
      "Python parent id": 2373, "Python id": 2385, "finished": true, "Ev Idx": 9215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564838663.189, "dur": 10.936,
    "args": {
      "Python parent id": 2373, "Python id": 2386, "finished": true, "Ev Idx": 9216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564838663.345, "dur": 0.072,
    "args": {
      "Python parent id": 2386, "Python id": 2387, "finished": true, "Ev Idx": 9217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838669.276, "dur": 1.211,
    "args": {
      "Python parent id": 2386, "Python id": 2388, "finished": true, "Ev Idx": 9218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564838670.021, "dur": 0.367,
    "args": {
      "Python parent id": 2388, "Python id": 2389, "finished": true, "Ev Idx": 9219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564838670.889, "dur": 3.145,
    "args": {
      "Python parent id": 2386, "Python id": 2390, "finished": true, "Ev Idx": 9220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564838674.833, "dur": 41.316,
    "args": {
      "Python parent id": 2373, "Python id": 2391, "finished": true, "Ev Idx": 9221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564838716.763, "dur": 2.925,
    "args": {
      "Python parent id": 2373, "Python id": 2392, "finished": true, "Ev Idx": 9222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564838720.377, "dur": 14.382,
    "args": {
      "Python parent id": 2373, "Python id": 2393, "finished": true, "Ev Idx": 9223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564838738.882, "dur": 3.384,
    "args": {
      "Python parent id": 2324, "Python id": 2394, "finished": true, "Ev Idx": 9224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564838742.901, "dur": 0.282,
    "args": {
      "Python parent id": 2324, "Python id": 2395, "finished": true, "Ev Idx": 9225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838744.172, "dur": 0.757,
    "args": {
      "Python parent id": 2324, "Python id": 2396, "finished": true, "Ev Idx": 9226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_94", "pid": 537, "tid": 537,
    "ts": 7429564838748.793, "dur": 42.849,
    "args": {
      "Python parent id": 2324, "Python module id": 94, "Python id": 2397, "finished": true, "Ev Idx": 9227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564838749.651, "dur": 41.644,
    "args": {
      "Python parent id": 2397, "Python id": 2398, "finished": true, "Ev Idx": 9228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564838750.034, "dur": 0.949,
    "args": {
      "Python parent id": 2398, "Python id": 2399, "finished": true, "Ev Idx": 9229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564838752.045, "dur": 39.007,
    "args": {
      "Python parent id": 2398, "Python id": 2400, "finished": true, "Ev Idx": 9230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838752.462, "dur": 0.225,
    "args": {
      "Python parent id": 2400, "Python id": 2401, "finished": true, "Ev Idx": 9231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838752.903, "dur": 0.144,
    "args": {
      "Python parent id": 2400, "Python id": 2402, "finished": true, "Ev Idx": 9232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564838753.207, "dur": 37.744,
    "args": {
      "Python parent id": 2400, "Python id": 2403, "finished": true, "Ev Idx": 9233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838806.180, "dur": 0.633,
    "args": {
      "Python parent id": 2308, "Python id": 2404, "finished": true, "Ev Idx": 9234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_27", "pid": 537, "tid": 537,
    "ts": 7429564838809.361, "dur": 86.244,
    "args": {
      "Python parent id": 2308, "Python module id": 27, "Python id": 2405, "finished": true, "Ev Idx": 9235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564838810.211, "dur": 85.059,
    "args": {
      "Python parent id": 2405, "Python id": 2406, "finished": true, "Ev Idx": 9236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564838810.511, "dur": 0.476,
    "args": {
      "Python parent id": 2406, "Python id": 2407, "finished": true, "Ev Idx": 9237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564838812.444, "dur": 82.006,
    "args": {
      "Python parent id": 2406, "Python id": 2408, "finished": true, "Ev Idx": 9238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564838812.889, "dur": 1.840,
    "args": {
      "Python parent id": 2408, "Python id": 2409, "finished": true, "Ev Idx": 9239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564838815.038, "dur": 13.209,
    "args": {
      "Python parent id": 2408, "Python id": 2410, "finished": true, "Ev Idx": 9240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564838828.633, "dur": 16.885,
    "args": {
      "Python parent id": 2408, "Python id": 2411, "finished": true, "Ev Idx": 9241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564838857.611, "dur": 12.606,
    "args": {
      "Python parent id": 2408, "Python id": 2412, "finished": true, "Ev Idx": 9242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838881.582, "dur": 0.367,
    "args": {
      "Python parent id": 2408, "Python id": 2413, "finished": true, "Ev Idx": 9243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564838882.362, "dur": 1.261,
    "args": {
      "Python parent id": 2408, "Python id": 2414, "finished": true, "Ev Idx": 9244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838896.356, "dur": 0.414,
    "args": {
      "Python parent id": 2308, "Python id": 2415, "finished": true, "Ev Idx": 9245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_13", "pid": 537, "tid": 537,
    "ts": 7429564838898.568, "dur": 274.885,
    "args": {
      "Python parent id": 2308, "Python module id": 13, "Python id": 2416, "finished": true, "Ev Idx": 9246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564838899.197, "dur": 274.017,
    "args": {
      "Python parent id": 2416, "Python id": 2417, "finished": true, "Ev Idx": 9247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564838899.395, "dur": 0.309,
    "args": {
      "Python parent id": 2417, "Python id": 2418, "finished": true, "Ev Idx": 9248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564838900.911, "dur": 272.172,
    "args": {
      "Python parent id": 2417, "Python id": 2419, "finished": true, "Ev Idx": 9249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838901.212, "dur": 0.326,
    "args": {
      "Python parent id": 2419, "Python id": 2420, "finished": true, "Ev Idx": 9250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838901.885, "dur": 0.303,
    "args": {
      "Python parent id": 2419, "Python id": 2421, "finished": true, "Ev Idx": 9251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838902.468, "dur": 0.232,
    "args": {
      "Python parent id": 2419, "Python id": 2422, "finished": true, "Ev Idx": 9252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_95", "pid": 537, "tid": 537,
    "ts": 7429564838904.394, "dur": 109.694,
    "args": {
      "Python parent id": 2419, "Python module id": 95, "Python id": 2423, "finished": true, "Ev Idx": 9253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564838904.856, "dur": 108.941,
    "args": {
      "Python parent id": 2423, "Python id": 2424, "finished": true, "Ev Idx": 9254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564838905.003, "dur": 0.167,
    "args": {
      "Python parent id": 2424, "Python id": 2425, "finished": true, "Ev Idx": 9255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564838905.878, "dur": 107.778,
    "args": {
      "Python parent id": 2424, "Python id": 2426, "finished": true, "Ev Idx": 9256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838906.173, "dur": 0.173,
    "args": {
      "Python parent id": 2426, "Python id": 2427, "finished": true, "Ev Idx": 9257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564838906.560, "dur": 0.132,
    "args": {
      "Python parent id": 2426, "Python id": 2428, "finished": true, "Ev Idx": 9258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564838906.832, "dur": 106.700,
    "args": {
      "Python parent id": 2426, "Python id": 2429, "finished": true, "Ev Idx": 9259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_13", "pid": 537, "tid": 537,
    "ts": 7429564839016.294, "dur": 17.063,
    "args": {
      "Python parent id": 2419, "Python module id": 13, "Python id": 2430, "finished": true, "Ev Idx": 9260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564839016.905, "dur": 16.262,
    "args": {
      "Python parent id": 2430, "Python id": 2431, "finished": true, "Ev Idx": 9261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564839017.219, "dur": 0.501,
    "args": {
      "Python parent id": 2431, "Python id": 2432, "finished": true, "Ev Idx": 9262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564839018.764, "dur": 14.303,
    "args": {
      "Python parent id": 2431, "Python id": 2433, "finished": true, "Ev Idx": 9263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564839019.571, "dur": 13.316,
    "args": {
      "Python parent id": 2433, "Python id": 2434, "finished": true, "Ev Idx": 9264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564839019.711, "dur": 0.230,
    "args": {
      "Python parent id": 2434, "Python id": 2435, "finished": true, "Ev Idx": 9265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564839020.792, "dur": 12.012,
    "args": {
      "Python parent id": 2434, "Python id": 2436, "finished": true, "Ev Idx": 9266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839034.392, "dur": 0.384,
    "args": {
      "Python parent id": 2419, "Python id": 2437, "finished": true, "Ev Idx": 9267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_96", "pid": 537, "tid": 537,
    "ts": 7429564839036.642, "dur": 79.753,
    "args": {
      "Python parent id": 2419, "Python module id": 96, "Python id": 2438, "finished": true, "Ev Idx": 9268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564839037.196, "dur": 78.910,
    "args": {
      "Python parent id": 2438, "Python id": 2439, "finished": true, "Ev Idx": 9269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564839037.399, "dur": 0.224,
    "args": {
      "Python parent id": 2439, "Python id": 2440, "finished": true, "Ev Idx": 9270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564839038.445, "dur": 77.538,
    "args": {
      "Python parent id": 2439, "Python id": 2441, "finished": true, "Ev Idx": 9271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839038.766, "dur": 0.181,
    "args": {
      "Python parent id": 2441, "Python id": 2442, "finished": true, "Ev Idx": 9272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839039.162, "dur": 0.146,
    "args": {
      "Python parent id": 2441, "Python id": 2443, "finished": true, "Ev Idx": 9273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564839039.416, "dur": 76.509,
    "args": {
      "Python parent id": 2441, "Python id": 2444, "finished": true, "Ev Idx": 9274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_97", "pid": 537, "tid": 537,
    "ts": 7429564839129.483, "dur": 43.303,
    "args": {
      "Python parent id": 2419, "Python module id": 97, "Python id": 2445, "finished": true, "Ev Idx": 9275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564839130.382, "dur": 42.099,
    "args": {
      "Python parent id": 2445, "Python id": 2446, "finished": true, "Ev Idx": 9276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564839130.637, "dur": 0.425,
    "args": {
      "Python parent id": 2446, "Python id": 2447, "finished": true, "Ev Idx": 9277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564839131.772, "dur": 40.575,
    "args": {
      "Python parent id": 2446, "Python id": 2448, "finished": true, "Ev Idx": 9278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839132.076, "dur": 0.166,
    "args": {
      "Python parent id": 2448, "Python id": 2449, "finished": true, "Ev Idx": 9279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839132.428, "dur": 0.120,
    "args": {
      "Python parent id": 2448, "Python id": 2450, "finished": true, "Ev Idx": 9280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564839132.660, "dur": 39.598,
    "args": {
      "Python parent id": 2448, "Python id": 2451, "finished": true, "Ev Idx": 9281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_14", "pid": 537, "tid": 537,
    "ts": 7429564839193.916, "dur": 1223.632,
    "args": {
      "Python parent id": 64, "Python module id": 14, "Python id": 2452, "finished": true, "Ev Idx": 9282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564839195.190, "dur": 1221.452,
    "args": {
      "Python parent id": 2452, "Python id": 2453, "finished": true, "Ev Idx": 9283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564839195.383, "dur": 0.361,
    "args": {
      "Python parent id": 2453, "Python id": 2454, "finished": true, "Ev Idx": 9284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564839197.428, "dur": 1218.264,
    "args": {
      "Python parent id": 2453, "Python id": 2455, "finished": true, "Ev Idx": 9285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839197.936, "dur": 0.414,
    "args": {
      "Python parent id": 2455, "Python id": 2456, "finished": true, "Ev Idx": 9286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_28", "pid": 537, "tid": 537,
    "ts": 7429564839200.382, "dur": 75.105,
    "args": {
      "Python parent id": 2455, "Python module id": 28, "Python id": 2457, "finished": true, "Ev Idx": 9287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564839200.887, "dur": 74.242,
    "args": {
      "Python parent id": 2457, "Python id": 2458, "finished": true, "Ev Idx": 9288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564839201.024, "dur": 0.149,
    "args": {
      "Python parent id": 2458, "Python id": 2459, "finished": true, "Ev Idx": 9289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564839201.847, "dur": 72.584,
    "args": {
      "Python parent id": 2458, "Python id": 2460, "finished": true, "Ev Idx": 9290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564839202.351, "dur": 1.797,
    "args": {
      "Python parent id": 2460, "Python id": 2461, "finished": true, "Ev Idx": 9291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564839204.402, "dur": 11.616,
    "args": {
      "Python parent id": 2460, "Python id": 2462, "finished": true, "Ev Idx": 9292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564839216.665, "dur": 14.477,
    "args": {
      "Python parent id": 2460, "Python id": 2463, "finished": true, "Ev Idx": 9293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564839242.597, "dur": 9.301,
    "args": {
      "Python parent id": 2460, "Python id": 2464, "finished": true, "Ev Idx": 9294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839262.722, "dur": 0.352,
    "args": {
      "Python parent id": 2460, "Python id": 2465, "finished": true, "Ev Idx": 9295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564839263.546, "dur": 1.445,
    "args": {
      "Python parent id": 2460, "Python id": 2466, "finished": true, "Ev Idx": 9296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839276.051, "dur": 0.319,
    "args": {
      "Python parent id": 2455, "Python id": 2467, "finished": true, "Ev Idx": 9297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_14", "pid": 537, "tid": 537,
    "ts": 7429564839278.979, "dur": 742.308,
    "args": {
      "Python parent id": 2455, "Python module id": 14, "Python id": 2468, "finished": true, "Ev Idx": 9298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564839280.442, "dur": 740.188,
    "args": {
      "Python parent id": 2468, "Python id": 2469, "finished": true, "Ev Idx": 9299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564839280.603, "dur": 0.270,
    "args": {
      "Python parent id": 2469, "Python id": 2470, "finished": true, "Ev Idx": 9300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564839282.339, "dur": 737.490,
    "args": {
      "Python parent id": 2469, "Python id": 2471, "finished": true, "Ev Idx": 9301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839285.297, "dur": 0.382,
    "args": {
      "Python parent id": 2471, "Python id": 2472, "finished": true, "Ev Idx": 9302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_98", "pid": 537, "tid": 537,
    "ts": 7429564839287.344, "dur": 41.794,
    "args": {
      "Python parent id": 2471, "Python module id": 98, "Python id": 2473, "finished": true, "Ev Idx": 9303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564839287.803, "dur": 41.058,
    "args": {
      "Python parent id": 2473, "Python id": 2474, "finished": true, "Ev Idx": 9304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564839287.947, "dur": 0.156,
    "args": {
      "Python parent id": 2474, "Python id": 2475, "finished": true, "Ev Idx": 9305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564839288.883, "dur": 39.865,
    "args": {
      "Python parent id": 2474, "Python id": 2476, "finished": true, "Ev Idx": 9306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839289.160, "dur": 0.286,
    "args": {
      "Python parent id": 2476, "Python id": 2477, "finished": true, "Ev Idx": 9307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839289.657, "dur": 0.115,
    "args": {
      "Python parent id": 2476, "Python id": 2478, "finished": true, "Ev Idx": 9308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564839289.878, "dur": 38.789,
    "args": {
      "Python parent id": 2476, "Python id": 2479, "finished": true, "Ev Idx": 9309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564839329.931, "dur": 3.202,
    "args": {
      "Python parent id": 2471, "Python id": 2480, "finished": true, "Ev Idx": 9310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564839333.724, "dur": 4.050,
    "args": {
      "Python parent id": 2471, "Python id": 2481, "finished": true, "Ev Idx": 9311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839338.638, "dur": 0.345,
    "args": {
      "Python parent id": 2471, "Python id": 2482, "finished": true, "Ev Idx": 9312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_99", "pid": 537, "tid": 537,
    "ts": 7429564839341.287, "dur": 77.488,
    "args": {
      "Python parent id": 2471, "Python module id": 99, "Python id": 2483, "finished": true, "Ev Idx": 9313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564839342.100, "dur": 76.456,
    "args": {
      "Python parent id": 2483, "Python id": 2484, "finished": true, "Ev Idx": 9314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564839342.325, "dur": 0.263,
    "args": {
      "Python parent id": 2484, "Python id": 2485, "finished": true, "Ev Idx": 9315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564839343.534, "dur": 74.918,
    "args": {
      "Python parent id": 2484, "Python id": 2486, "finished": true, "Ev Idx": 9316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839343.783, "dur": 0.148,
    "args": {
      "Python parent id": 2486, "Python id": 2487, "finished": true, "Ev Idx": 9317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839344.102, "dur": 0.127,
    "args": {
      "Python parent id": 2486, "Python id": 2488, "finished": true, "Ev Idx": 9318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564839344.363, "dur": 74.023,
    "args": {
      "Python parent id": 2486, "Python id": 2489, "finished": true, "Ev Idx": 9319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564839419.335, "dur": 2.471,
    "args": {
      "Python parent id": 2471, "Python id": 2490, "finished": true, "Ev Idx": 9320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564839422.346, "dur": 2.516,
    "args": {
      "Python parent id": 2471, "Python id": 2491, "finished": true, "Ev Idx": 9321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839425.722, "dur": 0.515,
    "args": {
      "Python parent id": 2471, "Python id": 2492, "finished": true, "Ev Idx": 9322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_100", "pid": 537, "tid": 537,
    "ts": 7429564839428.644, "dur": 40.071,
    "args": {
      "Python parent id": 2471, "Python module id": 100, "Python id": 2493, "finished": true, "Ev Idx": 9323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564839429.297, "dur": 39.111,
    "args": {
      "Python parent id": 2493, "Python id": 2494, "finished": true, "Ev Idx": 9324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564839429.537, "dur": 0.279,
    "args": {
      "Python parent id": 2494, "Python id": 2495, "finished": true, "Ev Idx": 9325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564839430.415, "dur": 37.875,
    "args": {
      "Python parent id": 2494, "Python id": 2496, "finished": true, "Ev Idx": 9326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839430.647, "dur": 0.141,
    "args": {
      "Python parent id": 2496, "Python id": 2497, "finished": true, "Ev Idx": 9327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839430.968, "dur": 0.105,
    "args": {
      "Python parent id": 2496, "Python id": 2498, "finished": true, "Ev Idx": 9328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564839431.192, "dur": 37.051,
    "args": {
      "Python parent id": 2496, "Python id": 2499, "finished": true, "Ev Idx": 9329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564839469.256, "dur": 2.426,
    "args": {
      "Python parent id": 2471, "Python id": 2500, "finished": true, "Ev Idx": 9330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564839472.214, "dur": 2.691,
    "args": {
      "Python parent id": 2471, "Python id": 2501, "finished": true, "Ev Idx": 9331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564839476.018, "dur": 198.839,
    "args": {
      "Python parent id": 2471, "Python id": 2502, "finished": true, "Ev Idx": 9332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564839476.334, "dur": 3.684,
    "args": {
      "Python parent id": 2502, "Python id": 2503, "finished": true, "Ev Idx": 9333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564839480.342, "dur": 1.879,
    "args": {
      "Python parent id": 2502, "Python id": 2504, "finished": true, "Ev Idx": 9334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564839494.092, "dur": 49.466,
    "args": {
      "Python parent id": 2502, "Python id": 2505, "finished": true, "Ev Idx": 9335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564839520.935, "dur": 21.451,
    "args": {
      "Python parent id": 2505, "Python id": 2506, "finished": true, "Ev Idx": 9336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564839623.370, "dur": 31.572,
    "args": {
      "Python parent id": 2502, "Python id": 2507, "finished": true, "Ev Idx": 9337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564839641.575, "dur": 12.605,
    "args": {
      "Python parent id": 2507, "Python id": 2508, "finished": true, "Ev Idx": 9338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564839679.727, "dur": 4.045,
    "args": {
      "Python parent id": 2471, "Python id": 2509, "finished": true, "Ev Idx": 9339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564839680.594, "dur": 0.377,
    "args": {
      "Python parent id": 2509, "Python id": 2510, "finished": true, "Ev Idx": 9340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564839681.313, "dur": 0.067,
    "args": {
      "Python parent id": 2509, "Python id": 2511, "finished": true, "Ev Idx": 9341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564839682.771, "dur": 0.224,
    "args": {
      "Python parent id": 2509, "Python id": 2512, "finished": true, "Ev Idx": 9342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564839683.416, "dur": 0.129,
    "args": {
      "Python parent id": 2509, "Python id": 2513, "finished": true, "Ev Idx": 9343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564839685.177, "dur": 6.746,
    "args": {
      "Python parent id": 2471, "Python id": 2514, "finished": true, "Ev Idx": 9344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564839688.383, "dur": 3.409,
    "args": {
      "Python parent id": 2514, "Python id": 2515, "finished": true, "Ev Idx": 9345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564839688.625, "dur": 1.522,
    "args": {
      "Python parent id": 2515, "Python id": 2516, "finished": true, "Ev Idx": 9346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564839689.083, "dur": 0.887,
    "args": {
      "Python parent id": 2516, "Python id": 2517, "finished": true, "Ev Idx": 9347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564839690.461, "dur": 0.476,
    "args": {
      "Python parent id": 2515, "Python id": 2518, "finished": true, "Ev Idx": 9348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564839691.264, "dur": 0.446,
    "args": {
      "Python parent id": 2515, "Python id": 2519, "finished": true, "Ev Idx": 9349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564839694.981, "dur": 264.054,
    "args": {
      "Python parent id": 2471, "Python id": 2520, "finished": true, "Ev Idx": 9350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564839695.598, "dur": 48.277,
    "args": {
      "Python parent id": 2520, "Python id": 2521, "finished": true, "Ev Idx": 9351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564839715.652, "dur": 5.563,
    "args": {
      "Python parent id": 2521, "Python id": 2522, "finished": true, "Ev Idx": 9352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564839722.104, "dur": 21.616,
    "args": {
      "Python parent id": 2521, "Python id": 2523, "finished": true, "Ev Idx": 9353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564839744.708, "dur": 32.242,
    "args": {
      "Python parent id": 2520, "Python id": 2524, "finished": true, "Ev Idx": 9354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564839759.096, "dur": 2.918,
    "args": {
      "Python parent id": 2524, "Python id": 2525, "finished": true, "Ev Idx": 9355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564839762.494, "dur": 14.389,
    "args": {
      "Python parent id": 2524, "Python id": 2526, "finished": true, "Ev Idx": 9356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564839777.947, "dur": 3.028,
    "args": {
      "Python parent id": 2520, "Python id": 2527, "finished": true, "Ev Idx": 9357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564839781.248, "dur": 49.468,
    "args": {
      "Python parent id": 2520, "Python id": 2528, "finished": true, "Ev Idx": 9358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564839865.078, "dur": 16.779,
    "args": {
      "Python parent id": 2520, "Python id": 2529, "finished": true, "Ev Idx": 9359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564839865.292, "dur": 0.319,
    "args": {
      "Python parent id": 2529, "Python id": 2530, "finished": true, "Ev Idx": 9360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564839866.233, "dur": 15.463,
    "args": {
      "Python parent id": 2529, "Python id": 2531, "finished": true, "Ev Idx": 9361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564839882.885, "dur": 2.051,
    "args": {
      "Python parent id": 2520, "Python id": 2532, "finished": true, "Ev Idx": 9362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564839886.396, "dur": 10.608,
    "args": {
      "Python parent id": 2520, "Python id": 2533, "finished": true, "Ev Idx": 9363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564839886.524, "dur": 0.145,
    "args": {
      "Python parent id": 2533, "Python id": 2534, "finished": true, "Ev Idx": 9364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839892.184, "dur": 1.435,
    "args": {
      "Python parent id": 2533, "Python id": 2535, "finished": true, "Ev Idx": 9365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564839893.052, "dur": 0.483,
    "args": {
      "Python parent id": 2535, "Python id": 2536, "finished": true, "Ev Idx": 9366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564839894.074, "dur": 2.855,
    "args": {
      "Python parent id": 2533, "Python id": 2537, "finished": true, "Ev Idx": 9367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564839897.795, "dur": 42.658,
    "args": {
      "Python parent id": 2520, "Python id": 2538, "finished": true, "Ev Idx": 9368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564839941.137, "dur": 3.013,
    "args": {
      "Python parent id": 2520, "Python id": 2539, "finished": true, "Ev Idx": 9369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564839944.825, "dur": 13.148,
    "args": {
      "Python parent id": 2520, "Python id": 2540, "finished": true, "Ev Idx": 9370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564839961.963, "dur": 3.498,
    "args": {
      "Python parent id": 2471, "Python id": 2541, "finished": true, "Ev Idx": 9371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564839966.137, "dur": 0.317,
    "args": {
      "Python parent id": 2471, "Python id": 2542, "finished": true, "Ev Idx": 9372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839967.437, "dur": 0.929,
    "args": {
      "Python parent id": 2471, "Python id": 2543, "finished": true, "Ev Idx": 9373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_101", "pid": 537, "tid": 537,
    "ts": 7429564839972.913, "dur": 46.345,
    "args": {
      "Python parent id": 2471, "Python module id": 101, "Python id": 2544, "finished": true, "Ev Idx": 9374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564839974.279, "dur": 44.517,
    "args": {
      "Python parent id": 2544, "Python id": 2545, "finished": true, "Ev Idx": 9375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564839974.672, "dur": 1.170,
    "args": {
      "Python parent id": 2545, "Python id": 2546, "finished": true, "Ev Idx": 9376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564839977.048, "dur": 41.530,
    "args": {
      "Python parent id": 2545, "Python id": 2547, "finished": true, "Ev Idx": 9377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839977.553, "dur": 0.250,
    "args": {
      "Python parent id": 2547, "Python id": 2548, "finished": true, "Ev Idx": 9378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564839978.117, "dur": 0.151,
    "args": {
      "Python parent id": 2547, "Python id": 2549, "finished": true, "Ev Idx": 9379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564839978.457, "dur": 40.016,
    "args": {
      "Python parent id": 2547, "Python id": 2550, "finished": true, "Ev Idx": 9380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840033.577, "dur": 0.584,
    "args": {
      "Python parent id": 2455, "Python id": 2551, "finished": true, "Ev Idx": 9381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_29", "pid": 537, "tid": 537,
    "ts": 7429564840036.873, "dur": 84.771,
    "args": {
      "Python parent id": 2455, "Python module id": 29, "Python id": 2552, "finished": true, "Ev Idx": 9382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840037.572, "dur": 83.733,
    "args": {
      "Python parent id": 2552, "Python id": 2553, "finished": true, "Ev Idx": 9383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840037.866, "dur": 0.432,
    "args": {
      "Python parent id": 2553, "Python id": 2554, "finished": true, "Ev Idx": 9384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564840039.155, "dur": 81.325,
    "args": {
      "Python parent id": 2553, "Python id": 2555, "finished": true, "Ev Idx": 9385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564840039.776, "dur": 1.753,
    "args": {
      "Python parent id": 2555, "Python id": 2556, "finished": true, "Ev Idx": 9386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564840041.886, "dur": 15.301,
    "args": {
      "Python parent id": 2555, "Python id": 2557, "finished": true, "Ev Idx": 9387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564840057.580, "dur": 15.558,
    "args": {
      "Python parent id": 2555, "Python id": 2558, "finished": true, "Ev Idx": 9388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564840086.228, "dur": 10.305,
    "args": {
      "Python parent id": 2555, "Python id": 2559, "finished": true, "Ev Idx": 9389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840108.325, "dur": 0.281,
    "args": {
      "Python parent id": 2555, "Python id": 2560, "finished": true, "Ev Idx": 9390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564840109.080, "dur": 1.647,
    "args": {
      "Python parent id": 2555, "Python id": 2561, "finished": true, "Ev Idx": 9391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840122.406, "dur": 0.398,
    "args": {
      "Python parent id": 2455, "Python id": 2562, "finished": true, "Ev Idx": 9392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_14", "pid": 537, "tid": 537,
    "ts": 7429564840124.655, "dur": 278.927,
    "args": {
      "Python parent id": 2455, "Python module id": 14, "Python id": 2563, "finished": true, "Ev Idx": 9393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840125.286, "dur": 278.106,
    "args": {
      "Python parent id": 2563, "Python id": 2564, "finished": true, "Ev Idx": 9394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840125.521, "dur": 0.349,
    "args": {
      "Python parent id": 2564, "Python id": 2565, "finished": true, "Ev Idx": 9395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564840127.301, "dur": 275.927,
    "args": {
      "Python parent id": 2564, "Python id": 2566, "finished": true, "Ev Idx": 9396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840127.629, "dur": 0.337,
    "args": {
      "Python parent id": 2566, "Python id": 2567, "finished": true, "Ev Idx": 9397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840128.356, "dur": 0.343,
    "args": {
      "Python parent id": 2566, "Python id": 2568, "finished": true, "Ev Idx": 9398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840128.990, "dur": 0.273,
    "args": {
      "Python parent id": 2566, "Python id": 2569, "finished": true, "Ev Idx": 9399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_102", "pid": 537, "tid": 537,
    "ts": 7429564840131.199, "dur": 53.608,
    "args": {
      "Python parent id": 2566, "Python module id": 102, "Python id": 2570, "finished": true, "Ev Idx": 9400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840131.815, "dur": 52.791,
    "args": {
      "Python parent id": 2570, "Python id": 2571, "finished": true, "Ev Idx": 9401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840131.969, "dur": 1.771,
    "args": {
      "Python parent id": 2571, "Python id": 2572, "finished": true, "Ev Idx": 9402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564840134.528, "dur": 49.859,
    "args": {
      "Python parent id": 2571, "Python id": 2573, "finished": true, "Ev Idx": 9403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840134.910, "dur": 0.140,
    "args": {
      "Python parent id": 2573, "Python id": 2574, "finished": true, "Ev Idx": 9404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840135.305, "dur": 0.158,
    "args": {
      "Python parent id": 2573, "Python id": 2575, "finished": true, "Ev Idx": 9405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564840135.611, "dur": 48.655,
    "args": {
      "Python parent id": 2573, "Python id": 2576, "finished": true, "Ev Idx": 9406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_14", "pid": 537, "tid": 537,
    "ts": 7429564840187.052, "dur": 66.403,
    "args": {
      "Python parent id": 2566, "Python module id": 14, "Python id": 2577, "finished": true, "Ev Idx": 9407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840187.733, "dur": 65.509,
    "args": {
      "Python parent id": 2577, "Python id": 2578, "finished": true, "Ev Idx": 9408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840187.944, "dur": 0.397,
    "args": {
      "Python parent id": 2578, "Python id": 2579, "finished": true, "Ev Idx": 9409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564840189.415, "dur": 63.737,
    "args": {
      "Python parent id": 2578, "Python id": 2580, "finished": true, "Ev Idx": 9410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564840190.398, "dur": 62.587,
    "args": {
      "Python parent id": 2580, "Python id": 2581, "finished": true, "Ev Idx": 9411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564840190.564, "dur": 0.204,
    "args": {
      "Python parent id": 2581, "Python id": 2582, "finished": true, "Ev Idx": 9412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564840191.733, "dur": 61.146,
    "args": {
      "Python parent id": 2581, "Python id": 2583, "finished": true, "Ev Idx": 9413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840254.423, "dur": 0.336,
    "args": {
      "Python parent id": 2566, "Python id": 2584, "finished": true, "Ev Idx": 9414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_103", "pid": 537, "tid": 537,
    "ts": 7429564840257.053, "dur": 50.372,
    "args": {
      "Python parent id": 2566, "Python module id": 103, "Python id": 2585, "finished": true, "Ev Idx": 9415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840257.570, "dur": 49.644,
    "args": {
      "Python parent id": 2585, "Python id": 2586, "finished": true, "Ev Idx": 9416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840257.795, "dur": 0.319,
    "args": {
      "Python parent id": 2586, "Python id": 2587, "finished": true, "Ev Idx": 9417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564840258.918, "dur": 48.162,
    "args": {
      "Python parent id": 2586, "Python id": 2588, "finished": true, "Ev Idx": 9418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840259.246, "dur": 0.124,
    "args": {
      "Python parent id": 2588, "Python id": 2589, "finished": true, "Ev Idx": 9419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840259.578, "dur": 0.148,
    "args": {
      "Python parent id": 2588, "Python id": 2590, "finished": true, "Ev Idx": 9420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564840259.850, "dur": 47.081,
    "args": {
      "Python parent id": 2588, "Python id": 2591, "finished": true, "Ev Idx": 9421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_104", "pid": 537, "tid": 537,
    "ts": 7429564840353.711, "dur": 49.254,
    "args": {
      "Python parent id": 2566, "Python module id": 104, "Python id": 2592, "finished": true, "Ev Idx": 9422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840354.508, "dur": 48.231,
    "args": {
      "Python parent id": 2592, "Python id": 2593, "finished": true, "Ev Idx": 9423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840354.776, "dur": 0.465,
    "args": {
      "Python parent id": 2593, "Python id": 2594, "finished": true, "Ev Idx": 9424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564840356.129, "dur": 46.433,
    "args": {
      "Python parent id": 2593, "Python id": 2595, "finished": true, "Ev Idx": 9425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840356.443, "dur": 0.186,
    "args": {
      "Python parent id": 2595, "Python id": 2596, "finished": true, "Ev Idx": 9426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840356.819, "dur": 0.147,
    "args": {
      "Python parent id": 2595, "Python id": 2597, "finished": true, "Ev Idx": 9427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564840357.075, "dur": 45.425,
    "args": {
      "Python parent id": 2595, "Python id": 2598, "finished": true, "Ev Idx": 9428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_15", "pid": 537, "tid": 537,
    "ts": 7429564840424.174, "dur": 1146.185,
    "args": {
      "Python parent id": 64, "Python module id": 15, "Python id": 2599, "finished": true, "Ev Idx": 9429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840425.588, "dur": 1143.962,
    "args": {
      "Python parent id": 2599, "Python id": 2600, "finished": true, "Ev Idx": 9430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840425.814, "dur": 0.358,
    "args": {
      "Python parent id": 2600, "Python id": 2601, "finished": true, "Ev Idx": 9431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564840427.745, "dur": 1140.854,
    "args": {
      "Python parent id": 2600, "Python id": 2602, "finished": true, "Ev Idx": 9432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840428.287, "dur": 0.362,
    "args": {
      "Python parent id": 2602, "Python id": 2603, "finished": true, "Ev Idx": 9433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_30", "pid": 537, "tid": 537,
    "ts": 7429564840431.140, "dur": 81.491,
    "args": {
      "Python parent id": 2602, "Python module id": 30, "Python id": 2604, "finished": true, "Ev Idx": 9434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840431.676, "dur": 80.654,
    "args": {
      "Python parent id": 2604, "Python id": 2605, "finished": true, "Ev Idx": 9435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840431.822, "dur": 0.173,
    "args": {
      "Python parent id": 2605, "Python id": 2606, "finished": true, "Ev Idx": 9436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564840432.746, "dur": 78.804,
    "args": {
      "Python parent id": 2605, "Python id": 2607, "finished": true, "Ev Idx": 9437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564840433.321, "dur": 1.962,
    "args": {
      "Python parent id": 2607, "Python id": 2608, "finished": true, "Ev Idx": 9438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564840435.615, "dur": 11.792,
    "args": {
      "Python parent id": 2607, "Python id": 2609, "finished": true, "Ev Idx": 9439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564840447.732, "dur": 17.127,
    "args": {
      "Python parent id": 2607, "Python id": 2610, "finished": true, "Ev Idx": 9440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564840477.217, "dur": 12.545,
    "args": {
      "Python parent id": 2607, "Python id": 2611, "finished": true, "Ev Idx": 9441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840500.275, "dur": 0.179,
    "args": {
      "Python parent id": 2607, "Python id": 2612, "finished": true, "Ev Idx": 9442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564840500.777, "dur": 1.195,
    "args": {
      "Python parent id": 2607, "Python id": 2613, "finished": true, "Ev Idx": 9443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840513.176, "dur": 0.342,
    "args": {
      "Python parent id": 2602, "Python id": 2614, "finished": true, "Ev Idx": 9444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_15", "pid": 537, "tid": 537,
    "ts": 7429564840516.343, "dur": 651.469,
    "args": {
      "Python parent id": 2602, "Python module id": 15, "Python id": 2615, "finished": true, "Ev Idx": 9445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840517.983, "dur": 649.176,
    "args": {
      "Python parent id": 2615, "Python id": 2616, "finished": true, "Ev Idx": 9446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840518.199, "dur": 0.322,
    "args": {
      "Python parent id": 2616, "Python id": 2617, "finished": true, "Ev Idx": 9447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564840520.141, "dur": 646.261,
    "args": {
      "Python parent id": 2616, "Python id": 2618, "finished": true, "Ev Idx": 9448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840523.371, "dur": 0.397,
    "args": {
      "Python parent id": 2618, "Python id": 2619, "finished": true, "Ev Idx": 9449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_105", "pid": 537, "tid": 537,
    "ts": 7429564840525.580, "dur": 42.050,
    "args": {
      "Python parent id": 2618, "Python module id": 105, "Python id": 2620, "finished": true, "Ev Idx": 9450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840526.140, "dur": 41.244,
    "args": {
      "Python parent id": 2620, "Python id": 2621, "finished": true, "Ev Idx": 9451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840526.292, "dur": 0.160,
    "args": {
      "Python parent id": 2621, "Python id": 2622, "finished": true, "Ev Idx": 9452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564840527.131, "dur": 40.145,
    "args": {
      "Python parent id": 2621, "Python id": 2623, "finished": true, "Ev Idx": 9453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840527.416, "dur": 0.153,
    "args": {
      "Python parent id": 2623, "Python id": 2624, "finished": true, "Ev Idx": 9454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840527.747, "dur": 0.145,
    "args": {
      "Python parent id": 2623, "Python id": 2625, "finished": true, "Ev Idx": 9455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564840527.997, "dur": 39.183,
    "args": {
      "Python parent id": 2623, "Python id": 2626, "finished": true, "Ev Idx": 9456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564840568.162, "dur": 3.306,
    "args": {
      "Python parent id": 2618, "Python id": 2627, "finished": true, "Ev Idx": 9457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564840571.871, "dur": 3.524,
    "args": {
      "Python parent id": 2618, "Python id": 2628, "finished": true, "Ev Idx": 9458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840576.123, "dur": 0.333,
    "args": {
      "Python parent id": 2618, "Python id": 2629, "finished": true, "Ev Idx": 9459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_106", "pid": 537, "tid": 537,
    "ts": 7429564840578.623, "dur": 47.723,
    "args": {
      "Python parent id": 2618, "Python module id": 106, "Python id": 2630, "finished": true, "Ev Idx": 9460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840579.059, "dur": 47.052,
    "args": {
      "Python parent id": 2630, "Python id": 2631, "finished": true, "Ev Idx": 9461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840579.271, "dur": 0.224,
    "args": {
      "Python parent id": 2631, "Python id": 2632, "finished": true, "Ev Idx": 9462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564840580.147, "dur": 45.859,
    "args": {
      "Python parent id": 2631, "Python id": 2633, "finished": true, "Ev Idx": 9463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840580.349, "dur": 0.119,
    "args": {
      "Python parent id": 2633, "Python id": 2634, "finished": true, "Ev Idx": 9464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840580.678, "dur": 0.115,
    "args": {
      "Python parent id": 2633, "Python id": 2635, "finished": true, "Ev Idx": 9465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564840580.899, "dur": 45.049,
    "args": {
      "Python parent id": 2633, "Python id": 2636, "finished": true, "Ev Idx": 9466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564840626.780, "dur": 2.583,
    "args": {
      "Python parent id": 2618, "Python id": 2637, "finished": true, "Ev Idx": 9467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564840629.910, "dur": 2.539,
    "args": {
      "Python parent id": 2618, "Python id": 2638, "finished": true, "Ev Idx": 9468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840633.066, "dur": 0.453,
    "args": {
      "Python parent id": 2618, "Python id": 2639, "finished": true, "Ev Idx": 9469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_107", "pid": 537, "tid": 537,
    "ts": 7429564840635.595, "dur": 42.206,
    "args": {
      "Python parent id": 2618, "Python module id": 107, "Python id": 2640, "finished": true, "Ev Idx": 9470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564840636.081, "dur": 41.464,
    "args": {
      "Python parent id": 2640, "Python id": 2641, "finished": true, "Ev Idx": 9471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564840636.265, "dur": 0.259,
    "args": {
      "Python parent id": 2641, "Python id": 2642, "finished": true, "Ev Idx": 9472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564840637.210, "dur": 40.234,
    "args": {
      "Python parent id": 2641, "Python id": 2643, "finished": true, "Ev Idx": 9473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840637.417, "dur": 0.135,
    "args": {
      "Python parent id": 2643, "Python id": 2644, "finished": true, "Ev Idx": 9474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564840637.776, "dur": 0.141,
    "args": {
      "Python parent id": 2643, "Python id": 2645, "finished": true, "Ev Idx": 9475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564840638.026, "dur": 39.371,
    "args": {
      "Python parent id": 2643, "Python id": 2646, "finished": true, "Ev Idx": 9476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564840678.339, "dur": 2.344,
    "args": {
      "Python parent id": 2618, "Python id": 2647, "finished": true, "Ev Idx": 9477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564840681.235, "dur": 2.513,
    "args": {
      "Python parent id": 2618, "Python id": 2648, "finished": true, "Ev Idx": 9478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564840684.794, "dur": 151.361,
    "args": {
      "Python parent id": 2618, "Python id": 2649, "finished": true, "Ev Idx": 9479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564840685.058, "dur": 3.832,
    "args": {
      "Python parent id": 2649, "Python id": 2650, "finished": true, "Ev Idx": 9480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564840689.140, "dur": 1.914,
    "args": {
      "Python parent id": 2649, "Python id": 2651, "finished": true, "Ev Idx": 9481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564840703.628, "dur": 51.353,
    "args": {
      "Python parent id": 2649, "Python id": 2652, "finished": true, "Ev Idx": 9482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564840733.859, "dur": 19.914,
    "args": {
      "Python parent id": 2652, "Python id": 2653, "finished": true, "Ev Idx": 9483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564840786.058, "dur": 32.907,
    "args": {
      "Python parent id": 2649, "Python id": 2654, "finished": true, "Ev Idx": 9484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564840803.955, "dur": 14.168,
    "args": {
      "Python parent id": 2654, "Python id": 2655, "finished": true, "Ev Idx": 9485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564840840.307, "dur": 3.855,
    "args": {
      "Python parent id": 2618, "Python id": 2656, "finished": true, "Ev Idx": 9486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564840841.077, "dur": 0.267,
    "args": {
      "Python parent id": 2656, "Python id": 2657, "finished": true, "Ev Idx": 9487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564840841.674, "dur": 0.087,
    "args": {
      "Python parent id": 2656, "Python id": 2658, "finished": true, "Ev Idx": 9488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564840843.077, "dur": 0.263,
    "args": {
      "Python parent id": 2656, "Python id": 2659, "finished": true, "Ev Idx": 9489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564840843.766, "dur": 0.142,
    "args": {
      "Python parent id": 2656, "Python id": 2660, "finished": true, "Ev Idx": 9490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564840845.513, "dur": 6.151,
    "args": {
      "Python parent id": 2618, "Python id": 2661, "finished": true, "Ev Idx": 9491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564840848.209, "dur": 3.326,
    "args": {
      "Python parent id": 2661, "Python id": 2662, "finished": true, "Ev Idx": 9492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564840848.446, "dur": 1.443,
    "args": {
      "Python parent id": 2662, "Python id": 2663, "finished": true, "Ev Idx": 9493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564840849.002, "dur": 0.737,
    "args": {
      "Python parent id": 2663, "Python id": 2664, "finished": true, "Ev Idx": 9494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564840850.148, "dur": 0.474,
    "args": {
      "Python parent id": 2662, "Python id": 2665, "finished": true, "Ev Idx": 9495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564840850.973, "dur": 0.476,
    "args": {
      "Python parent id": 2662, "Python id": 2666, "finished": true, "Ev Idx": 9496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564840854.085, "dur": 255.644,
    "args": {
      "Python parent id": 2618, "Python id": 2667, "finished": true, "Ev Idx": 9497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564840854.706, "dur": 45.934,
    "args": {
      "Python parent id": 2667, "Python id": 2668, "finished": true, "Ev Idx": 9498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564840872.521, "dur": 5.259,
    "args": {
      "Python parent id": 2668, "Python id": 2669, "finished": true, "Ev Idx": 9499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564840878.702, "dur": 21.789,
    "args": {
      "Python parent id": 2668, "Python id": 2670, "finished": true, "Ev Idx": 9500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564840901.505, "dur": 31.450,
    "args": {
      "Python parent id": 2667, "Python id": 2671, "finished": true, "Ev Idx": 9501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564840916.104, "dur": 3.403,
    "args": {
      "Python parent id": 2671, "Python id": 2672, "finished": true, "Ev Idx": 9502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564840920.014, "dur": 12.812,
    "args": {
      "Python parent id": 2671, "Python id": 2673, "finished": true, "Ev Idx": 9503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564840933.715, "dur": 2.896,
    "args": {
      "Python parent id": 2667, "Python id": 2674, "finished": true, "Ev Idx": 9504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564840936.919, "dur": 44.256,
    "args": {
      "Python parent id": 2667, "Python id": 2675, "finished": true, "Ev Idx": 9505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564841014.336, "dur": 16.627,
    "args": {
      "Python parent id": 2667, "Python id": 2676, "finished": true, "Ev Idx": 9506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564841014.563, "dur": 0.288,
    "args": {
      "Python parent id": 2676, "Python id": 2677, "finished": true, "Ev Idx": 9507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564841015.567, "dur": 15.273,
    "args": {
      "Python parent id": 2676, "Python id": 2678, "finished": true, "Ev Idx": 9508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564841031.978, "dur": 1.832,
    "args": {
      "Python parent id": 2667, "Python id": 2679, "finished": true, "Ev Idx": 9509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564841035.365, "dur": 10.198,
    "args": {
      "Python parent id": 2667, "Python id": 2680, "finished": true, "Ev Idx": 9510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564841035.489, "dur": 0.150,
    "args": {
      "Python parent id": 2680, "Python id": 2681, "finished": true, "Ev Idx": 9511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841040.944, "dur": 1.292,
    "args": {
      "Python parent id": 2680, "Python id": 2682, "finished": true, "Ev Idx": 9512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564841041.755, "dur": 0.397,
    "args": {
      "Python parent id": 2682, "Python id": 2683, "finished": true, "Ev Idx": 9513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564841042.682, "dur": 2.794,
    "args": {
      "Python parent id": 2680, "Python id": 2684, "finished": true, "Ev Idx": 9514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564841046.262, "dur": 43.804,
    "args": {
      "Python parent id": 2667, "Python id": 2685, "finished": true, "Ev Idx": 9515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564841090.765, "dur": 3.143,
    "args": {
      "Python parent id": 2667, "Python id": 2686, "finished": true, "Ev Idx": 9516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564841094.636, "dur": 13.983,
    "args": {
      "Python parent id": 2667, "Python id": 2687, "finished": true, "Ev Idx": 9517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564841112.645, "dur": 3.472,
    "args": {
      "Python parent id": 2618, "Python id": 2688, "finished": true, "Ev Idx": 9518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564841117.013, "dur": 0.332,
    "args": {
      "Python parent id": 2618, "Python id": 2689, "finished": true, "Ev Idx": 9519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841118.489, "dur": 0.977,
    "args": {
      "Python parent id": 2618, "Python id": 2690, "finished": true, "Ev Idx": 9520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_108", "pid": 537, "tid": 537,
    "ts": 7429564841123.741, "dur": 42.045,
    "args": {
      "Python parent id": 2618, "Python module id": 108, "Python id": 2691, "finished": true, "Ev Idx": 9521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841124.747, "dur": 40.524,
    "args": {
      "Python parent id": 2691, "Python id": 2692, "finished": true, "Ev Idx": 9522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841125.239, "dur": 1.000,
    "args": {
      "Python parent id": 2692, "Python id": 2693, "finished": true, "Ev Idx": 9523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564841127.399, "dur": 37.575,
    "args": {
      "Python parent id": 2692, "Python id": 2694, "finished": true, "Ev Idx": 9524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841127.856, "dur": 0.266,
    "args": {
      "Python parent id": 2694, "Python id": 2695, "finished": true, "Ev Idx": 9525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841128.511, "dur": 0.179,
    "args": {
      "Python parent id": 2694, "Python id": 2696, "finished": true, "Ev Idx": 9526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564841128.879, "dur": 35.967,
    "args": {
      "Python parent id": 2694, "Python id": 2697, "finished": true, "Ev Idx": 9527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841180.345, "dur": 0.588,
    "args": {
      "Python parent id": 2602, "Python id": 2698, "finished": true, "Ev Idx": 9528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_31", "pid": 537, "tid": 537,
    "ts": 7429564841183.621, "dur": 81.091,
    "args": {
      "Python parent id": 2602, "Python module id": 31, "Python id": 2699, "finished": true, "Ev Idx": 9529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841184.389, "dur": 79.974,
    "args": {
      "Python parent id": 2699, "Python id": 2700, "finished": true, "Ev Idx": 9530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841184.644, "dur": 0.423,
    "args": {
      "Python parent id": 2700, "Python id": 2701, "finished": true, "Ev Idx": 9531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564841185.886, "dur": 77.576,
    "args": {
      "Python parent id": 2700, "Python id": 2702, "finished": true, "Ev Idx": 9532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564841186.471, "dur": 1.824,
    "args": {
      "Python parent id": 2702, "Python id": 2703, "finished": true, "Ev Idx": 9533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564841188.618, "dur": 13.202,
    "args": {
      "Python parent id": 2702, "Python id": 2704, "finished": true, "Ev Idx": 9534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564841202.188, "dur": 13.982,
    "args": {
      "Python parent id": 2702, "Python id": 2705, "finished": true, "Ev Idx": 9535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564841230.224, "dur": 11.142,
    "args": {
      "Python parent id": 2702, "Python id": 2706, "finished": true, "Ev Idx": 9536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841251.278, "dur": 0.256,
    "args": {
      "Python parent id": 2702, "Python id": 2707, "finished": true, "Ev Idx": 9537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564841252.034, "dur": 1.475,
    "args": {
      "Python parent id": 2702, "Python id": 2708, "finished": true, "Ev Idx": 9538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841265.473, "dur": 0.441,
    "args": {
      "Python parent id": 2602, "Python id": 2709, "finished": true, "Ev Idx": 9539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_15", "pid": 537, "tid": 537,
    "ts": 7429564841267.793, "dur": 288.692,
    "args": {
      "Python parent id": 2602, "Python module id": 15, "Python id": 2710, "finished": true, "Ev Idx": 9540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841268.466, "dur": 287.823,
    "args": {
      "Python parent id": 2710, "Python id": 2711, "finished": true, "Ev Idx": 9541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841268.668, "dur": 0.378,
    "args": {
      "Python parent id": 2711, "Python id": 2712, "finished": true, "Ev Idx": 9542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564841270.253, "dur": 285.856,
    "args": {
      "Python parent id": 2711, "Python id": 2713, "finished": true, "Ev Idx": 9543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841270.520, "dur": 0.355,
    "args": {
      "Python parent id": 2713, "Python id": 2714, "finished": true, "Ev Idx": 9544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841271.266, "dur": 0.273,
    "args": {
      "Python parent id": 2713, "Python id": 2715, "finished": true, "Ev Idx": 9545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841271.796, "dur": 0.237,
    "args": {
      "Python parent id": 2713, "Python id": 2716, "finished": true, "Ev Idx": 9546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_109", "pid": 537, "tid": 537,
    "ts": 7429564841274.254, "dur": 115.125,
    "args": {
      "Python parent id": 2713, "Python module id": 109, "Python id": 2717, "finished": true, "Ev Idx": 9547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841274.867, "dur": 114.248,
    "args": {
      "Python parent id": 2717, "Python id": 2718, "finished": true, "Ev Idx": 9548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841275.016, "dur": 0.139,
    "args": {
      "Python parent id": 2718, "Python id": 2719, "finished": true, "Ev Idx": 9549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564841275.905, "dur": 112.997,
    "args": {
      "Python parent id": 2718, "Python id": 2720, "finished": true, "Ev Idx": 9550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841276.152, "dur": 0.150,
    "args": {
      "Python parent id": 2720, "Python id": 2721, "finished": true, "Ev Idx": 9551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841276.528, "dur": 0.131,
    "args": {
      "Python parent id": 2720, "Python id": 2722, "finished": true, "Ev Idx": 9552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564841276.787, "dur": 112.012,
    "args": {
      "Python parent id": 2720, "Python id": 2723, "finished": true, "Ev Idx": 9553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_15", "pid": 537, "tid": 537,
    "ts": 7429564841391.410, "dur": 19.753,
    "args": {
      "Python parent id": 2713, "Python module id": 15, "Python id": 2724, "finished": true, "Ev Idx": 9554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841391.988, "dur": 18.854,
    "args": {
      "Python parent id": 2724, "Python id": 2725, "finished": true, "Ev Idx": 9555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841392.219, "dur": 0.425,
    "args": {
      "Python parent id": 2725, "Python id": 2726, "finished": true, "Ev Idx": 9556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564841393.818, "dur": 16.909,
    "args": {
      "Python parent id": 2725, "Python id": 2727, "finished": true, "Ev Idx": 9557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564841394.760, "dur": 15.783,
    "args": {
      "Python parent id": 2727, "Python id": 2728, "finished": true, "Ev Idx": 9558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564841394.920, "dur": 0.250,
    "args": {
      "Python parent id": 2728, "Python id": 2729, "finished": true, "Ev Idx": 9559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564841396.061, "dur": 14.372,
    "args": {
      "Python parent id": 2728, "Python id": 2730, "finished": true, "Ev Idx": 9560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841411.929, "dur": 0.452,
    "args": {
      "Python parent id": 2713, "Python id": 2731, "finished": true, "Ev Idx": 9561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_110", "pid": 537, "tid": 537,
    "ts": 7429564841414.365, "dur": 84.210,
    "args": {
      "Python parent id": 2713, "Python module id": 110, "Python id": 2732, "finished": true, "Ev Idx": 9562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841414.813, "dur": 83.515,
    "args": {
      "Python parent id": 2732, "Python id": 2733, "finished": true, "Ev Idx": 9563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841415.043, "dur": 0.229,
    "args": {
      "Python parent id": 2733, "Python id": 2734, "finished": true, "Ev Idx": 9564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564841415.866, "dur": 82.274,
    "args": {
      "Python parent id": 2733, "Python id": 2735, "finished": true, "Ev Idx": 9565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841416.115, "dur": 0.186,
    "args": {
      "Python parent id": 2735, "Python id": 2736, "finished": true, "Ev Idx": 9566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841416.534, "dur": 0.164,
    "args": {
      "Python parent id": 2735, "Python id": 2737, "finished": true, "Ev Idx": 9567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564841416.818, "dur": 81.264,
    "args": {
      "Python parent id": 2735, "Python id": 2738, "finished": true, "Ev Idx": 9568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_111", "pid": 537, "tid": 537,
    "ts": 7429564841512.135, "dur": 43.642,
    "args": {
      "Python parent id": 2713, "Python module id": 111, "Python id": 2739, "finished": true, "Ev Idx": 9569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841512.847, "dur": 42.627,
    "args": {
      "Python parent id": 2739, "Python id": 2740, "finished": true, "Ev Idx": 9570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841513.136, "dur": 0.332,
    "args": {
      "Python parent id": 2740, "Python id": 2741, "finished": true, "Ev Idx": 9571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564841514.354, "dur": 40.974,
    "args": {
      "Python parent id": 2740, "Python id": 2742, "finished": true, "Ev Idx": 9572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841514.678, "dur": 0.312,
    "args": {
      "Python parent id": 2742, "Python id": 2743, "finished": true, "Ev Idx": 9573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841515.319, "dur": 0.128,
    "args": {
      "Python parent id": 2742, "Python id": 2744, "finished": true, "Ev Idx": 9574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564841515.558, "dur": 39.722,
    "args": {
      "Python parent id": 2742, "Python id": 2745, "finished": true, "Ev Idx": 9575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_16", "pid": 537, "tid": 537,
    "ts": 7429564841577.159, "dur": 1253.792,
    "args": {
      "Python parent id": 64, "Python module id": 16, "Python id": 2746, "finished": true, "Ev Idx": 9576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841578.806, "dur": 1251.328,
    "args": {
      "Python parent id": 2746, "Python id": 2747, "finished": true, "Ev Idx": 9577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841579.046, "dur": 0.397,
    "args": {
      "Python parent id": 2747, "Python id": 2748, "finished": true, "Ev Idx": 9578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564841581.051, "dur": 1248.194,
    "args": {
      "Python parent id": 2747, "Python id": 2749, "finished": true, "Ev Idx": 9579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841581.484, "dur": 0.540,
    "args": {
      "Python parent id": 2749, "Python id": 2750, "finished": true, "Ev Idx": 9580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_32", "pid": 537, "tid": 537,
    "ts": 7429564841583.939, "dur": 76.658,
    "args": {
      "Python parent id": 2749, "Python module id": 32, "Python id": 2751, "finished": true, "Ev Idx": 9581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841584.513, "dur": 75.713,
    "args": {
      "Python parent id": 2751, "Python id": 2752, "finished": true, "Ev Idx": 9582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841584.662, "dur": 0.176,
    "args": {
      "Python parent id": 2752, "Python id": 2753, "finished": true, "Ev Idx": 9583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564841585.665, "dur": 73.838,
    "args": {
      "Python parent id": 2752, "Python id": 2754, "finished": true, "Ev Idx": 9584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564841586.110, "dur": 1.853,
    "args": {
      "Python parent id": 2754, "Python id": 2755, "finished": true, "Ev Idx": 9585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564841588.235, "dur": 11.974,
    "args": {
      "Python parent id": 2754, "Python id": 2756, "finished": true, "Ev Idx": 9586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564841600.687, "dur": 13.340,
    "args": {
      "Python parent id": 2754, "Python id": 2757, "finished": true, "Ev Idx": 9587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564841626.262, "dur": 10.918,
    "args": {
      "Python parent id": 2754, "Python id": 2758, "finished": true, "Ev Idx": 9588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841646.550, "dur": 0.198,
    "args": {
      "Python parent id": 2754, "Python id": 2759, "finished": true, "Ev Idx": 9589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564841647.148, "dur": 1.220,
    "args": {
      "Python parent id": 2754, "Python id": 2760, "finished": true, "Ev Idx": 9590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841661.234, "dur": 0.358,
    "args": {
      "Python parent id": 2749, "Python id": 2761, "finished": true, "Ev Idx": 9591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_16", "pid": 537, "tid": 537,
    "ts": 7429564841664.492, "dur": 769.107,
    "args": {
      "Python parent id": 2749, "Python module id": 16, "Python id": 2762, "finished": true, "Ev Idx": 9592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841666.070, "dur": 766.779,
    "args": {
      "Python parent id": 2762, "Python id": 2763, "finished": true, "Ev Idx": 9593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841666.266, "dur": 0.279,
    "args": {
      "Python parent id": 2763, "Python id": 2764, "finished": true, "Ev Idx": 9594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564841668.245, "dur": 763.780,
    "args": {
      "Python parent id": 2763, "Python id": 2765, "finished": true, "Ev Idx": 9595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841671.619, "dur": 0.311,
    "args": {
      "Python parent id": 2765, "Python id": 2766, "finished": true, "Ev Idx": 9596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_112", "pid": 537, "tid": 537,
    "ts": 7429564841674.218, "dur": 41.189,
    "args": {
      "Python parent id": 2765, "Python module id": 112, "Python id": 2767, "finished": true, "Ev Idx": 9597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841674.902, "dur": 40.199,
    "args": {
      "Python parent id": 2767, "Python id": 2768, "finished": true, "Ev Idx": 9598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841675.059, "dur": 0.340,
    "args": {
      "Python parent id": 2768, "Python id": 2769, "finished": true, "Ev Idx": 9599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564841676.125, "dur": 38.875,
    "args": {
      "Python parent id": 2768, "Python id": 2770, "finished": true, "Ev Idx": 9600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841676.412, "dur": 0.128,
    "args": {
      "Python parent id": 2770, "Python id": 2771, "finished": true, "Ev Idx": 9601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841676.795, "dur": 0.128,
    "args": {
      "Python parent id": 2770, "Python id": 2772, "finished": true, "Ev Idx": 9602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564841677.041, "dur": 37.869,
    "args": {
      "Python parent id": 2770, "Python id": 2773, "finished": true, "Ev Idx": 9603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564841716.098, "dur": 3.184,
    "args": {
      "Python parent id": 2765, "Python id": 2774, "finished": true, "Ev Idx": 9604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564841719.778, "dur": 3.952,
    "args": {
      "Python parent id": 2765, "Python id": 2775, "finished": true, "Ev Idx": 9605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841724.556, "dur": 0.336,
    "args": {
      "Python parent id": 2765, "Python id": 2776, "finished": true, "Ev Idx": 9606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_113", "pid": 537, "tid": 537,
    "ts": 7429564841727.075, "dur": 43.205,
    "args": {
      "Python parent id": 2765, "Python module id": 113, "Python id": 2777, "finished": true, "Ev Idx": 9607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841727.589, "dur": 42.504,
    "args": {
      "Python parent id": 2777, "Python id": 2778, "finished": true, "Ev Idx": 9608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841727.815, "dur": 0.248,
    "args": {
      "Python parent id": 2778, "Python id": 2779, "finished": true, "Ev Idx": 9609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564841728.657, "dur": 41.333,
    "args": {
      "Python parent id": 2778, "Python id": 2780, "finished": true, "Ev Idx": 9610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841728.886, "dur": 0.140,
    "args": {
      "Python parent id": 2780, "Python id": 2781, "finished": true, "Ev Idx": 9611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841729.197, "dur": 0.114,
    "args": {
      "Python parent id": 2780, "Python id": 2782, "finished": true, "Ev Idx": 9612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564841729.418, "dur": 40.515,
    "args": {
      "Python parent id": 2780, "Python id": 2783, "finished": true, "Ev Idx": 9613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564841770.868, "dur": 2.528,
    "args": {
      "Python parent id": 2765, "Python id": 2784, "finished": true, "Ev Idx": 9614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564841773.863, "dur": 2.980,
    "args": {
      "Python parent id": 2765, "Python id": 2785, "finished": true, "Ev Idx": 9615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841777.575, "dur": 0.436,
    "args": {
      "Python parent id": 2765, "Python id": 2786, "finished": true, "Ev Idx": 9616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_114", "pid": 537, "tid": 537,
    "ts": 7429564841779.804, "dur": 39.102,
    "args": {
      "Python parent id": 2765, "Python module id": 114, "Python id": 2787, "finished": true, "Ev Idx": 9617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564841780.246, "dur": 38.495,
    "args": {
      "Python parent id": 2787, "Python id": 2788, "finished": true, "Ev Idx": 9618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564841780.476, "dur": 0.321,
    "args": {
      "Python parent id": 2788, "Python id": 2789, "finished": true, "Ev Idx": 9619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564841781.464, "dur": 37.048,
    "args": {
      "Python parent id": 2788, "Python id": 2790, "finished": true, "Ev Idx": 9620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841781.650, "dur": 0.125,
    "args": {
      "Python parent id": 2790, "Python id": 2791, "finished": true, "Ev Idx": 9621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564841781.944, "dur": 0.122,
    "args": {
      "Python parent id": 2790, "Python id": 2792, "finished": true, "Ev Idx": 9622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564841782.171, "dur": 36.289,
    "args": {
      "Python parent id": 2790, "Python id": 2793, "finished": true, "Ev Idx": 9623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564841819.457, "dur": 2.653,
    "args": {
      "Python parent id": 2765, "Python id": 2794, "finished": true, "Ev Idx": 9624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564841822.645, "dur": 2.469,
    "args": {
      "Python parent id": 2765, "Python id": 2795, "finished": true, "Ev Idx": 9625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564841826.282, "dur": 188.006,
    "args": {
      "Python parent id": 2765, "Python id": 2796, "finished": true, "Ev Idx": 9626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564841826.571, "dur": 3.519,
    "args": {
      "Python parent id": 2796, "Python id": 2797, "finished": true, "Ev Idx": 9627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564841830.373, "dur": 1.788,
    "args": {
      "Python parent id": 2796, "Python id": 2798, "finished": true, "Ev Idx": 9628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564841880.628, "dur": 53.234,
    "args": {
      "Python parent id": 2796, "Python id": 2799, "finished": true, "Ev Idx": 9629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564841911.046, "dur": 21.607,
    "args": {
      "Python parent id": 2799, "Python id": 2800, "finished": true, "Ev Idx": 9630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564841964.293, "dur": 32.628,
    "args": {
      "Python parent id": 2796, "Python id": 2801, "finished": true, "Ev Idx": 9631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564841982.103, "dur": 13.905,
    "args": {
      "Python parent id": 2801, "Python id": 2802, "finished": true, "Ev Idx": 9632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564842018.855, "dur": 4.110,
    "args": {
      "Python parent id": 2765, "Python id": 2803, "finished": true, "Ev Idx": 9633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564842019.672, "dur": 0.332,
    "args": {
      "Python parent id": 2803, "Python id": 2804, "finished": true, "Ev Idx": 9634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564842020.351, "dur": 0.087,
    "args": {
      "Python parent id": 2803, "Python id": 2805, "finished": true, "Ev Idx": 9635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564842021.702, "dur": 0.461,
    "args": {
      "Python parent id": 2803, "Python id": 2806, "finished": true, "Ev Idx": 9636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564842022.551, "dur": 0.150,
    "args": {
      "Python parent id": 2803, "Python id": 2807, "finished": true, "Ev Idx": 9637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564842024.303, "dur": 6.740,
    "args": {
      "Python parent id": 2765, "Python id": 2808, "finished": true, "Ev Idx": 9638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564842027.343, "dur": 3.572,
    "args": {
      "Python parent id": 2808, "Python id": 2809, "finished": true, "Ev Idx": 9639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564842027.552, "dur": 1.610,
    "args": {
      "Python parent id": 2809, "Python id": 2810, "finished": true, "Ev Idx": 9640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564842028.153, "dur": 0.834,
    "args": {
      "Python parent id": 2810, "Python id": 2811, "finished": true, "Ev Idx": 9641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564842029.478, "dur": 0.463,
    "args": {
      "Python parent id": 2809, "Python id": 2812, "finished": true, "Ev Idx": 9642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564842030.397, "dur": 0.446,
    "args": {
      "Python parent id": 2809, "Python id": 2813, "finished": true, "Ev Idx": 9643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564842033.769, "dur": 337.885,
    "args": {
      "Python parent id": 2765, "Python id": 2814, "finished": true, "Ev Idx": 9644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564842034.321, "dur": 72.443,
    "args": {
      "Python parent id": 2814, "Python id": 2815, "finished": true, "Ev Idx": 9645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564842053.475, "dur": 5.190,
    "args": {
      "Python parent id": 2815, "Python id": 2816, "finished": true, "Ev Idx": 9646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564842059.519, "dur": 47.085,
    "args": {
      "Python parent id": 2815, "Python id": 2817, "finished": true, "Ev Idx": 9647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564842107.992, "dur": 35.607,
    "args": {
      "Python parent id": 2814, "Python id": 2818, "finished": true, "Ev Idx": 9648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564842124.902, "dur": 3.502,
    "args": {
      "Python parent id": 2818, "Python id": 2819, "finished": true, "Ev Idx": 9649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564842128.957, "dur": 14.532,
    "args": {
      "Python parent id": 2818, "Python id": 2820, "finished": true, "Ev Idx": 9650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564842144.469, "dur": 3.259,
    "args": {
      "Python parent id": 2814, "Python id": 2821, "finished": true, "Ev Idx": 9651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564842148.048, "dur": 56.775,
    "args": {
      "Python parent id": 2814, "Python id": 2822, "finished": true, "Ev Idx": 9652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564842278.735, "dur": 18.087,
    "args": {
      "Python parent id": 2814, "Python id": 2823, "finished": true, "Ev Idx": 9653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564842278.934, "dur": 0.278,
    "args": {
      "Python parent id": 2823, "Python id": 2824, "finished": true, "Ev Idx": 9654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564842279.873, "dur": 16.808,
    "args": {
      "Python parent id": 2823, "Python id": 2825, "finished": true, "Ev Idx": 9655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564842297.728, "dur": 2.019,
    "args": {
      "Python parent id": 2814, "Python id": 2826, "finished": true, "Ev Idx": 9656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564842301.272, "dur": 9.821,
    "args": {
      "Python parent id": 2814, "Python id": 2827, "finished": true, "Ev Idx": 9657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564842301.356, "dur": 0.056,
    "args": {
      "Python parent id": 2827, "Python id": 2828, "finished": true, "Ev Idx": 9658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842306.992, "dur": 0.936,
    "args": {
      "Python parent id": 2827, "Python id": 2829, "finished": true, "Ev Idx": 9659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564842307.446, "dur": 0.407,
    "args": {
      "Python parent id": 2829, "Python id": 2830, "finished": true, "Ev Idx": 9660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564842308.277, "dur": 2.747,
    "args": {
      "Python parent id": 2827, "Python id": 2831, "finished": true, "Ev Idx": 9661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564842311.670, "dur": 41.841,
    "args": {
      "Python parent id": 2814, "Python id": 2832, "finished": true, "Ev Idx": 9662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564842354.075, "dur": 3.067,
    "args": {
      "Python parent id": 2814, "Python id": 2833, "finished": true, "Ev Idx": 9663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564842357.820, "dur": 12.816,
    "args": {
      "Python parent id": 2814, "Python id": 2834, "finished": true, "Ev Idx": 9664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564842374.865, "dur": 3.213,
    "args": {
      "Python parent id": 2765, "Python id": 2835, "finished": true, "Ev Idx": 9665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564842378.771, "dur": 0.287,
    "args": {
      "Python parent id": 2765, "Python id": 2836, "finished": true, "Ev Idx": 9666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842380.193, "dur": 1.210,
    "args": {
      "Python parent id": 2765, "Python id": 2837, "finished": true, "Ev Idx": 9667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_115", "pid": 537, "tid": 537,
    "ts": 7429564842386.432, "dur": 44.915,
    "args": {
      "Python parent id": 2765, "Python module id": 115, "Python id": 2838, "finished": true, "Ev Idx": 9668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842387.584, "dur": 43.349,
    "args": {
      "Python parent id": 2838, "Python id": 2839, "finished": true, "Ev Idx": 9669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842388.219, "dur": 1.344,
    "args": {
      "Python parent id": 2839, "Python id": 2840, "finished": true, "Ev Idx": 9670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564842390.838, "dur": 39.774,
    "args": {
      "Python parent id": 2839, "Python id": 2841, "finished": true, "Ev Idx": 9671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842391.409, "dur": 0.220,
    "args": {
      "Python parent id": 2841, "Python id": 2842, "finished": true, "Ev Idx": 9672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842392.027, "dur": 0.157,
    "args": {
      "Python parent id": 2841, "Python id": 2843, "finished": true, "Ev Idx": 9673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564842392.359, "dur": 38.168,
    "args": {
      "Python parent id": 2841, "Python id": 2844, "finished": true, "Ev Idx": 9674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842445.507, "dur": 0.559,
    "args": {
      "Python parent id": 2749, "Python id": 2845, "finished": true, "Ev Idx": 9675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_33", "pid": 537, "tid": 537,
    "ts": 7429564842449.090, "dur": 81.333,
    "args": {
      "Python parent id": 2749, "Python module id": 33, "Python id": 2846, "finished": true, "Ev Idx": 9676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842449.755, "dur": 80.289,
    "args": {
      "Python parent id": 2846, "Python id": 2847, "finished": true, "Ev Idx": 9677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842449.956, "dur": 0.473,
    "args": {
      "Python parent id": 2847, "Python id": 2848, "finished": true, "Ev Idx": 9678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564842451.378, "dur": 77.855,
    "args": {
      "Python parent id": 2847, "Python id": 2849, "finished": true, "Ev Idx": 9679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564842451.875, "dur": 1.460,
    "args": {
      "Python parent id": 2849, "Python id": 2850, "finished": true, "Ev Idx": 9680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564842453.616, "dur": 12.623,
    "args": {
      "Python parent id": 2849, "Python id": 2851, "finished": true, "Ev Idx": 9681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564842466.639, "dur": 15.275,
    "args": {
      "Python parent id": 2849, "Python id": 2852, "finished": true, "Ev Idx": 9682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564842494.173, "dur": 12.083,
    "args": {
      "Python parent id": 2849, "Python id": 2853, "finished": true, "Ev Idx": 9683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842517.887, "dur": 0.219,
    "args": {
      "Python parent id": 2849, "Python id": 2854, "finished": true, "Ev Idx": 9684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564842518.460, "dur": 1.219,
    "args": {
      "Python parent id": 2849, "Python id": 2855, "finished": true, "Ev Idx": 9685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842531.140, "dur": 0.448,
    "args": {
      "Python parent id": 2749, "Python id": 2856, "finished": true, "Ev Idx": 9686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_16", "pid": 537, "tid": 537,
    "ts": 7429564842533.104, "dur": 284.453,
    "args": {
      "Python parent id": 2749, "Python module id": 16, "Python id": 2857, "finished": true, "Ev Idx": 9687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842533.649, "dur": 283.649,
    "args": {
      "Python parent id": 2857, "Python id": 2858, "finished": true, "Ev Idx": 9688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842533.849, "dur": 0.346,
    "args": {
      "Python parent id": 2858, "Python id": 2859, "finished": true, "Ev Idx": 9689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564842535.383, "dur": 281.672,
    "args": {
      "Python parent id": 2858, "Python id": 2860, "finished": true, "Ev Idx": 9690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842535.675, "dur": 0.304,
    "args": {
      "Python parent id": 2860, "Python id": 2861, "finished": true, "Ev Idx": 9691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842536.311, "dur": 0.289,
    "args": {
      "Python parent id": 2860, "Python id": 2862, "finished": true, "Ev Idx": 9692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842536.828, "dur": 0.335,
    "args": {
      "Python parent id": 2860, "Python id": 2863, "finished": true, "Ev Idx": 9693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_116", "pid": 537, "tid": 537,
    "ts": 7429564842539.007, "dur": 52.314,
    "args": {
      "Python parent id": 2860, "Python module id": 116, "Python id": 2864, "finished": true, "Ev Idx": 9694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842539.541, "dur": 51.564,
    "args": {
      "Python parent id": 2864, "Python id": 2865, "finished": true, "Ev Idx": 9695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842539.678, "dur": 0.170,
    "args": {
      "Python parent id": 2865, "Python id": 2866, "finished": true, "Ev Idx": 9696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564842540.450, "dur": 50.536,
    "args": {
      "Python parent id": 2865, "Python id": 2867, "finished": true, "Ev Idx": 9697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842540.764, "dur": 0.121,
    "args": {
      "Python parent id": 2867, "Python id": 2868, "finished": true, "Ev Idx": 9698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842541.133, "dur": 0.140,
    "args": {
      "Python parent id": 2867, "Python id": 2869, "finished": true, "Ev Idx": 9699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564842541.427, "dur": 49.446,
    "args": {
      "Python parent id": 2867, "Python id": 2870, "finished": true, "Ev Idx": 9700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_16", "pid": 537, "tid": 537,
    "ts": 7429564842593.340, "dur": 77.630,
    "args": {
      "Python parent id": 2860, "Python module id": 16, "Python id": 2871, "finished": true, "Ev Idx": 9701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842593.918, "dur": 76.812,
    "args": {
      "Python parent id": 2871, "Python id": 2872, "finished": true, "Ev Idx": 9702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842594.134, "dur": 0.333,
    "args": {
      "Python parent id": 2872, "Python id": 2873, "finished": true, "Ev Idx": 9703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564842595.538, "dur": 75.056,
    "args": {
      "Python parent id": 2872, "Python id": 2874, "finished": true, "Ev Idx": 9704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564842596.659, "dur": 73.736,
    "args": {
      "Python parent id": 2874, "Python id": 2875, "finished": true, "Ev Idx": 9705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564842596.861, "dur": 0.199,
    "args": {
      "Python parent id": 2875, "Python id": 2876, "finished": true, "Ev Idx": 9706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564842597.982, "dur": 72.245,
    "args": {
      "Python parent id": 2875, "Python id": 2877, "finished": true, "Ev Idx": 9707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842671.919, "dur": 0.515,
    "args": {
      "Python parent id": 2860, "Python id": 2878, "finished": true, "Ev Idx": 9708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_117", "pid": 537, "tid": 537,
    "ts": 7429564842674.447, "dur": 50.275,
    "args": {
      "Python parent id": 2860, "Python module id": 117, "Python id": 2879, "finished": true, "Ev Idx": 9709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842674.950, "dur": 49.530,
    "args": {
      "Python parent id": 2879, "Python id": 2880, "finished": true, "Ev Idx": 9710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842675.228, "dur": 0.325,
    "args": {
      "Python parent id": 2880, "Python id": 2881, "finished": true, "Ev Idx": 9711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564842676.340, "dur": 47.978,
    "args": {
      "Python parent id": 2880, "Python id": 2882, "finished": true, "Ev Idx": 9712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842676.611, "dur": 0.171,
    "args": {
      "Python parent id": 2882, "Python id": 2883, "finished": true, "Ev Idx": 9713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842677.044, "dur": 0.152,
    "args": {
      "Python parent id": 2882, "Python id": 2884, "finished": true, "Ev Idx": 9714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564842677.312, "dur": 46.905,
    "args": {
      "Python parent id": 2882, "Python id": 2885, "finished": true, "Ev Idx": 9715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_118", "pid": 537, "tid": 537,
    "ts": 7429564842771.385, "dur": 45.373,
    "args": {
      "Python parent id": 2860, "Python module id": 118, "Python id": 2886, "finished": true, "Ev Idx": 9716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842772.067, "dur": 44.355,
    "args": {
      "Python parent id": 2886, "Python id": 2887, "finished": true, "Ev Idx": 9717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842772.364, "dur": 0.331,
    "args": {
      "Python parent id": 2887, "Python id": 2888, "finished": true, "Ev Idx": 9718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564842773.492, "dur": 42.769,
    "args": {
      "Python parent id": 2887, "Python id": 2889, "finished": true, "Ev Idx": 9719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842773.774, "dur": 0.174,
    "args": {
      "Python parent id": 2889, "Python id": 2890, "finished": true, "Ev Idx": 9720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842774.178, "dur": 0.170,
    "args": {
      "Python parent id": 2889, "Python id": 2891, "finished": true, "Ev Idx": 9721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564842774.478, "dur": 41.685,
    "args": {
      "Python parent id": 2889, "Python id": 2892, "finished": true, "Ev Idx": 9722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_17", "pid": 537, "tid": 537,
    "ts": 7429564842837.474, "dur": 1160.836,
    "args": {
      "Python parent id": 64, "Python module id": 17, "Python id": 2893, "finished": true, "Ev Idx": 9723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842838.900, "dur": 1158.591,
    "args": {
      "Python parent id": 2893, "Python id": 2894, "finished": true, "Ev Idx": 9724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842839.194, "dur": 0.310,
    "args": {
      "Python parent id": 2894, "Python id": 2895, "finished": true, "Ev Idx": 9725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564842841.466, "dur": 1155.181,
    "args": {
      "Python parent id": 2894, "Python id": 2896, "finished": true, "Ev Idx": 9726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842841.939, "dur": 0.721,
    "args": {
      "Python parent id": 2896, "Python id": 2897, "finished": true, "Ev Idx": 9727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_34", "pid": 537, "tid": 537,
    "ts": 7429564842844.799, "dur": 77.578,
    "args": {
      "Python parent id": 2896, "Python module id": 34, "Python id": 2898, "finished": true, "Ev Idx": 9728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842845.413, "dur": 76.660,
    "args": {
      "Python parent id": 2898, "Python id": 2899, "finished": true, "Ev Idx": 9729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842845.611, "dur": 0.176,
    "args": {
      "Python parent id": 2899, "Python id": 2900, "finished": true, "Ev Idx": 9730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564842846.744, "dur": 74.605,
    "args": {
      "Python parent id": 2899, "Python id": 2901, "finished": true, "Ev Idx": 9731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564842847.367, "dur": 2.021,
    "args": {
      "Python parent id": 2901, "Python id": 2902, "finished": true, "Ev Idx": 9732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564842849.624, "dur": 13.213,
    "args": {
      "Python parent id": 2901, "Python id": 2903, "finished": true, "Ev Idx": 9733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564842863.240, "dur": 15.012,
    "args": {
      "Python parent id": 2901, "Python id": 2904, "finished": true, "Ev Idx": 9734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564842890.509, "dur": 10.506,
    "args": {
      "Python parent id": 2901, "Python id": 2905, "finished": true, "Ev Idx": 9735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842910.051, "dur": 0.219,
    "args": {
      "Python parent id": 2901, "Python id": 2906, "finished": true, "Ev Idx": 9736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564842910.775, "dur": 1.125,
    "args": {
      "Python parent id": 2901, "Python id": 2907, "finished": true, "Ev Idx": 9737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842923.004, "dur": 0.347,
    "args": {
      "Python parent id": 2896, "Python id": 2908, "finished": true, "Ev Idx": 9738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_17", "pid": 537, "tid": 537,
    "ts": 7429564842926.309, "dur": 651.112,
    "args": {
      "Python parent id": 2896, "Python module id": 17, "Python id": 2909, "finished": true, "Ev Idx": 9739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842927.811, "dur": 649.157,
    "args": {
      "Python parent id": 2909, "Python id": 2910, "finished": true, "Ev Idx": 9740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842927.982, "dur": 0.346,
    "args": {
      "Python parent id": 2910, "Python id": 2911, "finished": true, "Ev Idx": 9741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564842930.009, "dur": 646.133,
    "args": {
      "Python parent id": 2910, "Python id": 2912, "finished": true, "Ev Idx": 9742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842933.588, "dur": 0.351,
    "args": {
      "Python parent id": 2912, "Python id": 2913, "finished": true, "Ev Idx": 9743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_119", "pid": 537, "tid": 537,
    "ts": 7429564842935.834, "dur": 40.086,
    "args": {
      "Python parent id": 2912, "Python module id": 119, "Python id": 2914, "finished": true, "Ev Idx": 9744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842936.361, "dur": 39.299,
    "args": {
      "Python parent id": 2914, "Python id": 2915, "finished": true, "Ev Idx": 9745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842936.528, "dur": 0.150,
    "args": {
      "Python parent id": 2915, "Python id": 2916, "finished": true, "Ev Idx": 9746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564842937.520, "dur": 37.983,
    "args": {
      "Python parent id": 2915, "Python id": 2917, "finished": true, "Ev Idx": 9747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842937.944, "dur": 0.124,
    "args": {
      "Python parent id": 2917, "Python id": 2918, "finished": true, "Ev Idx": 9748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842938.258, "dur": 0.145,
    "args": {
      "Python parent id": 2917, "Python id": 2919, "finished": true, "Ev Idx": 9749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564842938.523, "dur": 36.878,
    "args": {
      "Python parent id": 2917, "Python id": 2920, "finished": true, "Ev Idx": 9750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564842976.573, "dur": 3.451,
    "args": {
      "Python parent id": 2912, "Python id": 2921, "finished": true, "Ev Idx": 9751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564842980.400, "dur": 3.464,
    "args": {
      "Python parent id": 2912, "Python id": 2922, "finished": true, "Ev Idx": 9752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842984.541, "dur": 0.329,
    "args": {
      "Python parent id": 2912, "Python id": 2923, "finished": true, "Ev Idx": 9753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_120", "pid": 537, "tid": 537,
    "ts": 7429564842987.102, "dur": 44.202,
    "args": {
      "Python parent id": 2912, "Python module id": 120, "Python id": 2924, "finished": true, "Ev Idx": 9754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564842987.836, "dur": 43.241,
    "args": {
      "Python parent id": 2924, "Python id": 2925, "finished": true, "Ev Idx": 9755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564842988.032, "dur": 0.309,
    "args": {
      "Python parent id": 2925, "Python id": 2926, "finished": true, "Ev Idx": 9756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564842989.148, "dur": 41.817,
    "args": {
      "Python parent id": 2925, "Python id": 2927, "finished": true, "Ev Idx": 9757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842989.408, "dur": 0.152,
    "args": {
      "Python parent id": 2927, "Python id": 2928, "finished": true, "Ev Idx": 9758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564842989.727, "dur": 0.165,
    "args": {
      "Python parent id": 2927, "Python id": 2929, "finished": true, "Ev Idx": 9759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564842990.005, "dur": 40.862,
    "args": {
      "Python parent id": 2927, "Python id": 2930, "finished": true, "Ev Idx": 9760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564843031.773, "dur": 2.632,
    "args": {
      "Python parent id": 2912, "Python id": 2931, "finished": true, "Ev Idx": 9761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564843034.834, "dur": 2.844,
    "args": {
      "Python parent id": 2912, "Python id": 2932, "finished": true, "Ev Idx": 9762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843038.340, "dur": 0.413,
    "args": {
      "Python parent id": 2912, "Python id": 2933, "finished": true, "Ev Idx": 9763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_121", "pid": 537, "tid": 537,
    "ts": 7429564843040.888, "dur": 41.329,
    "args": {
      "Python parent id": 2912, "Python module id": 121, "Python id": 2934, "finished": true, "Ev Idx": 9764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564843041.654, "dur": 40.363,
    "args": {
      "Python parent id": 2934, "Python id": 2935, "finished": true, "Ev Idx": 9765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564843041.886, "dur": 0.243,
    "args": {
      "Python parent id": 2935, "Python id": 2936, "finished": true, "Ev Idx": 9766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564843042.732, "dur": 39.194,
    "args": {
      "Python parent id": 2935, "Python id": 2937, "finished": true, "Ev Idx": 9767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843042.981, "dur": 0.150,
    "args": {
      "Python parent id": 2937, "Python id": 2938, "finished": true, "Ev Idx": 9768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843043.313, "dur": 0.113,
    "args": {
      "Python parent id": 2937, "Python id": 2939, "finished": true, "Ev Idx": 9769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564843043.530, "dur": 38.293,
    "args": {
      "Python parent id": 2937, "Python id": 2940, "finished": true, "Ev Idx": 9770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564843082.762, "dur": 2.711,
    "args": {
      "Python parent id": 2912, "Python id": 2941, "finished": true, "Ev Idx": 9771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564843086.034, "dur": 2.596,
    "args": {
      "Python parent id": 2912, "Python id": 2942, "finished": true, "Ev Idx": 9772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564843089.737, "dur": 150.017,
    "args": {
      "Python parent id": 2912, "Python id": 2943, "finished": true, "Ev Idx": 9773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564843089.978, "dur": 3.660,
    "args": {
      "Python parent id": 2943, "Python id": 2944, "finished": true, "Ev Idx": 9774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564843093.935, "dur": 2.586,
    "args": {
      "Python parent id": 2943, "Python id": 2945, "finished": true, "Ev Idx": 9775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564843108.259, "dur": 51.115,
    "args": {
      "Python parent id": 2943, "Python id": 2946, "finished": true, "Ev Idx": 9776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564843136.461, "dur": 21.768,
    "args": {
      "Python parent id": 2946, "Python id": 2947, "finished": true, "Ev Idx": 9777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564843187.935, "dur": 31.174,
    "args": {
      "Python parent id": 2943, "Python id": 2948, "finished": true, "Ev Idx": 9778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564843205.346, "dur": 12.918,
    "args": {
      "Python parent id": 2948, "Python id": 2949, "finished": true, "Ev Idx": 9779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564843244.148, "dur": 3.850,
    "args": {
      "Python parent id": 2912, "Python id": 2950, "finished": true, "Ev Idx": 9780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564843244.874, "dur": 0.309,
    "args": {
      "Python parent id": 2950, "Python id": 2951, "finished": true, "Ev Idx": 9781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564843245.571, "dur": 0.123,
    "args": {
      "Python parent id": 2950, "Python id": 2952, "finished": true, "Ev Idx": 9782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564843246.964, "dur": 0.272,
    "args": {
      "Python parent id": 2950, "Python id": 2953, "finished": true, "Ev Idx": 9783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564843247.636, "dur": 0.125,
    "args": {
      "Python parent id": 2950, "Python id": 2954, "finished": true, "Ev Idx": 9784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564843249.455, "dur": 6.675,
    "args": {
      "Python parent id": 2912, "Python id": 2955, "finished": true, "Ev Idx": 9785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564843252.691, "dur": 3.324,
    "args": {
      "Python parent id": 2955, "Python id": 2956, "finished": true, "Ev Idx": 9786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564843252.940, "dur": 1.443,
    "args": {
      "Python parent id": 2956, "Python id": 2957, "finished": true, "Ev Idx": 9787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564843253.519, "dur": 0.721,
    "args": {
      "Python parent id": 2957, "Python id": 2958, "finished": true, "Ev Idx": 9788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564843254.629, "dur": 0.525,
    "args": {
      "Python parent id": 2956, "Python id": 2959, "finished": true, "Ev Idx": 9789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564843255.503, "dur": 0.431,
    "args": {
      "Python parent id": 2956, "Python id": 2960, "finished": true, "Ev Idx": 9790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564843258.942, "dur": 259.283,
    "args": {
      "Python parent id": 2912, "Python id": 2961, "finished": true, "Ev Idx": 9791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564843259.524, "dur": 46.190,
    "args": {
      "Python parent id": 2961, "Python id": 2962, "finished": true, "Ev Idx": 9792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564843279.787, "dur": 5.053,
    "args": {
      "Python parent id": 2962, "Python id": 2963, "finished": true, "Ev Idx": 9793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564843285.752, "dur": 19.845,
    "args": {
      "Python parent id": 2962, "Python id": 2964, "finished": true, "Ev Idx": 9794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564843306.602, "dur": 34.084,
    "args": {
      "Python parent id": 2961, "Python id": 2965, "finished": true, "Ev Idx": 9795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564843321.776, "dur": 2.802,
    "args": {
      "Python parent id": 2965, "Python id": 2966, "finished": true, "Ev Idx": 9796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564843325.094, "dur": 15.525,
    "args": {
      "Python parent id": 2965, "Python id": 2967, "finished": true, "Ev Idx": 9797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564843341.685, "dur": 3.023,
    "args": {
      "Python parent id": 2961, "Python id": 2968, "finished": true, "Ev Idx": 9798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564843344.968, "dur": 49.596,
    "args": {
      "Python parent id": 2961, "Python id": 2969, "finished": true, "Ev Idx": 9799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564843426.961, "dur": 17.707,
    "args": {
      "Python parent id": 2961, "Python id": 2970, "finished": true, "Ev Idx": 9800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564843427.140, "dur": 0.279,
    "args": {
      "Python parent id": 2970, "Python id": 2971, "finished": true, "Ev Idx": 9801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564843428.002, "dur": 16.533,
    "args": {
      "Python parent id": 2970, "Python id": 2972, "finished": true, "Ev Idx": 9802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564843445.564, "dur": 1.803,
    "args": {
      "Python parent id": 2961, "Python id": 2973, "finished": true, "Ev Idx": 9803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564843448.847, "dur": 9.708,
    "args": {
      "Python parent id": 2961, "Python id": 2974, "finished": true, "Ev Idx": 9804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564843448.997, "dur": 0.057,
    "args": {
      "Python parent id": 2974, "Python id": 2975, "finished": true, "Ev Idx": 9805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843454.414, "dur": 0.907,
    "args": {
      "Python parent id": 2974, "Python id": 2976, "finished": true, "Ev Idx": 9806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564843454.867, "dur": 0.368,
    "args": {
      "Python parent id": 2976, "Python id": 2977, "finished": true, "Ev Idx": 9807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564843455.727, "dur": 2.746,
    "args": {
      "Python parent id": 2974, "Python id": 2978, "finished": true, "Ev Idx": 9808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564843459.185, "dur": 40.312,
    "args": {
      "Python parent id": 2961, "Python id": 2979, "finished": true, "Ev Idx": 9809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564843500.128, "dur": 3.034,
    "args": {
      "Python parent id": 2961, "Python id": 2980, "finished": true, "Ev Idx": 9810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564843503.887, "dur": 13.421,
    "args": {
      "Python parent id": 2961, "Python id": 2981, "finished": true, "Ev Idx": 9811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564843521.124, "dur": 3.241,
    "args": {
      "Python parent id": 2912, "Python id": 2982, "finished": true, "Ev Idx": 9812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564843525.074, "dur": 0.274,
    "args": {
      "Python parent id": 2912, "Python id": 2983, "finished": true, "Ev Idx": 9813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843526.284, "dur": 0.963,
    "args": {
      "Python parent id": 2912, "Python id": 2984, "finished": true, "Ev Idx": 9814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_122", "pid": 537, "tid": 537,
    "ts": 7429564843531.754, "dur": 43.748,
    "args": {
      "Python parent id": 2912, "Python module id": 122, "Python id": 2985, "finished": true, "Ev Idx": 9815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564843532.973, "dur": 42.222,
    "args": {
      "Python parent id": 2985, "Python id": 2986, "finished": true, "Ev Idx": 9816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564843533.407, "dur": 1.002,
    "args": {
      "Python parent id": 2986, "Python id": 2987, "finished": true, "Ev Idx": 9817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564843536.004, "dur": 38.890,
    "args": {
      "Python parent id": 2986, "Python id": 2988, "finished": true, "Ev Idx": 9818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843536.370, "dur": 0.258,
    "args": {
      "Python parent id": 2988, "Python id": 2989, "finished": true, "Ev Idx": 9819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843536.969, "dur": 0.212,
    "args": {
      "Python parent id": 2988, "Python id": 2990, "finished": true, "Ev Idx": 9820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564843537.377, "dur": 37.385,
    "args": {
      "Python parent id": 2988, "Python id": 2991, "finished": true, "Ev Idx": 9821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843646.612, "dur": 1.009,
    "args": {
      "Python parent id": 2896, "Python id": 2992, "finished": true, "Ev Idx": 9822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_35", "pid": 537, "tid": 537,
    "ts": 7429564843651.188, "dur": 90.329,
    "args": {
      "Python parent id": 2896, "Python module id": 35, "Python id": 2993, "finished": true, "Ev Idx": 9823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564843652.076, "dur": 89.071,
    "args": {
      "Python parent id": 2993, "Python id": 2994, "finished": true, "Ev Idx": 9824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564843652.453, "dur": 0.623,
    "args": {
      "Python parent id": 2994, "Python id": 2995, "finished": true, "Ev Idx": 9825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564843654.860, "dur": 85.358,
    "args": {
      "Python parent id": 2994, "Python id": 2996, "finished": true, "Ev Idx": 9826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564843655.454, "dur": 2.105,
    "args": {
      "Python parent id": 2996, "Python id": 2997, "finished": true, "Ev Idx": 9827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564843657.933, "dur": 15.071,
    "args": {
      "Python parent id": 2996, "Python id": 2998, "finished": true, "Ev Idx": 9828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564843673.451, "dur": 17.803,
    "args": {
      "Python parent id": 2996, "Python id": 2999, "finished": true, "Ev Idx": 9829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564843703.845, "dur": 12.063,
    "args": {
      "Python parent id": 2996, "Python id": 3000, "finished": true, "Ev Idx": 9830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843727.693, "dur": 0.281,
    "args": {
      "Python parent id": 2996, "Python id": 3001, "finished": true, "Ev Idx": 9831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564843728.349, "dur": 1.174,
    "args": {
      "Python parent id": 2996, "Python id": 3002, "finished": true, "Ev Idx": 9832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843742.360, "dur": 0.395,
    "args": {
      "Python parent id": 2896, "Python id": 3003, "finished": true, "Ev Idx": 9833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_17", "pid": 537, "tid": 537,
    "ts": 7429564843744.410, "dur": 241.460,
    "args": {
      "Python parent id": 2896, "Python module id": 17, "Python id": 3004, "finished": true, "Ev Idx": 9834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564843745.085, "dur": 240.462,
    "args": {
      "Python parent id": 3004, "Python id": 3005, "finished": true, "Ev Idx": 9835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564843745.381, "dur": 0.405,
    "args": {
      "Python parent id": 3005, "Python id": 3006, "finished": true, "Ev Idx": 9836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564843747.052, "dur": 238.342,
    "args": {
      "Python parent id": 3005, "Python id": 3007, "finished": true, "Ev Idx": 9837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843747.347, "dur": 0.317,
    "args": {
      "Python parent id": 3007, "Python id": 3008, "finished": true, "Ev Idx": 9838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843748.007, "dur": 0.343,
    "args": {
      "Python parent id": 3007, "Python id": 3009, "finished": true, "Ev Idx": 9839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843748.606, "dur": 0.319,
    "args": {
      "Python parent id": 3007, "Python id": 3010, "finished": true, "Ev Idx": 9840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_123", "pid": 537, "tid": 537,
    "ts": 7429564843750.634, "dur": 59.258,
    "args": {
      "Python parent id": 3007, "Python module id": 123, "Python id": 3011, "finished": true, "Ev Idx": 9841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564843751.321, "dur": 58.341,
    "args": {
      "Python parent id": 3011, "Python id": 3012, "finished": true, "Ev Idx": 9842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564843751.476, "dur": 0.197,
    "args": {
      "Python parent id": 3012, "Python id": 3013, "finished": true, "Ev Idx": 9843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564843752.579, "dur": 56.875,
    "args": {
      "Python parent id": 3012, "Python id": 3014, "finished": true, "Ev Idx": 9844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843752.899, "dur": 0.167,
    "args": {
      "Python parent id": 3014, "Python id": 3015, "finished": true, "Ev Idx": 9845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843753.369, "dur": 0.165,
    "args": {
      "Python parent id": 3014, "Python id": 3016, "finished": true, "Ev Idx": 9846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564843753.716, "dur": 55.616,
    "args": {
      "Python parent id": 3014, "Python id": 3017, "finished": true, "Ev Idx": 9847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_17", "pid": 537, "tid": 537,
    "ts": 7429564843811.649, "dur": 17.396,
    "args": {
      "Python parent id": 3007, "Python module id": 17, "Python id": 3018, "finished": true, "Ev Idx": 9848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564843812.154, "dur": 16.675,
    "args": {
      "Python parent id": 3018, "Python id": 3019, "finished": true, "Ev Idx": 9849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564843812.371, "dur": 0.321,
    "args": {
      "Python parent id": 3019, "Python id": 3020, "finished": true, "Ev Idx": 9850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564843813.759, "dur": 14.993,
    "args": {
      "Python parent id": 3019, "Python id": 3021, "finished": true, "Ev Idx": 9851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564843814.669, "dur": 13.920,
    "args": {
      "Python parent id": 3021, "Python id": 3022, "finished": true, "Ev Idx": 9852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564843814.848, "dur": 0.161,
    "args": {
      "Python parent id": 3022, "Python id": 3023, "finished": true, "Ev Idx": 9853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564843815.962, "dur": 12.531,
    "args": {
      "Python parent id": 3022, "Python id": 3024, "finished": true, "Ev Idx": 9854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843829.714, "dur": 0.536,
    "args": {
      "Python parent id": 3007, "Python id": 3025, "finished": true, "Ev Idx": 9855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_124", "pid": 537, "tid": 537,
    "ts": 7429564843832.285, "dur": 90.585,
    "args": {
      "Python parent id": 3007, "Python module id": 124, "Python id": 3026, "finished": true, "Ev Idx": 9856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564843832.743, "dur": 89.919,
    "args": {
      "Python parent id": 3026, "Python id": 3027, "finished": true, "Ev Idx": 9857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564843832.938, "dur": 0.234,
    "args": {
      "Python parent id": 3027, "Python id": 3028, "finished": true, "Ev Idx": 9858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564843834.070, "dur": 88.381,
    "args": {
      "Python parent id": 3027, "Python id": 3029, "finished": true, "Ev Idx": 9859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843834.305, "dur": 0.179,
    "args": {
      "Python parent id": 3029, "Python id": 3030, "finished": true, "Ev Idx": 9860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843834.703, "dur": 0.159,
    "args": {
      "Python parent id": 3029, "Python id": 3031, "finished": true, "Ev Idx": 9861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564843834.992, "dur": 87.362,
    "args": {
      "Python parent id": 3029, "Python id": 3032, "finished": true, "Ev Idx": 9862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_125", "pid": 537, "tid": 537,
    "ts": 7429564843936.322, "dur": 48.749,
    "args": {
      "Python parent id": 3007, "Python module id": 125, "Python id": 3033, "finished": true, "Ev Idx": 9863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564843937.018, "dur": 47.886,
    "args": {
      "Python parent id": 3033, "Python id": 3034, "finished": true, "Ev Idx": 9864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564843937.298, "dur": 0.355,
    "args": {
      "Python parent id": 3034, "Python id": 3035, "finished": true, "Ev Idx": 9865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564843938.515, "dur": 46.242,
    "args": {
      "Python parent id": 3034, "Python id": 3036, "finished": true, "Ev Idx": 9866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843938.856, "dur": 0.214,
    "args": {
      "Python parent id": 3036, "Python id": 3037, "finished": true, "Ev Idx": 9867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564843939.314, "dur": 0.129,
    "args": {
      "Python parent id": 3036, "Python id": 3038, "finished": true, "Ev Idx": 9868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564843939.565, "dur": 45.107,
    "args": {
      "Python parent id": 3036, "Python id": 3039, "finished": true, "Ev Idx": 9869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_18", "pid": 537, "tid": 537,
    "ts": 7429564844004.868, "dur": 1206.728,
    "args": {
      "Python parent id": 64, "Python module id": 18, "Python id": 3040, "finished": true, "Ev Idx": 9870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564844006.425, "dur": 1204.489,
    "args": {
      "Python parent id": 3040, "Python id": 3041, "finished": true, "Ev Idx": 9871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564844006.596, "dur": 0.357,
    "args": {
      "Python parent id": 3041, "Python id": 3042, "finished": true, "Ev Idx": 9872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564844008.692, "dur": 1201.459,
    "args": {
      "Python parent id": 3041, "Python id": 3043, "finished": true, "Ev Idx": 9873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844009.211, "dur": 0.677,
    "args": {
      "Python parent id": 3043, "Python id": 3044, "finished": true, "Ev Idx": 9874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_36", "pid": 537, "tid": 537,
    "ts": 7429564844011.922, "dur": 126.246,
    "args": {
      "Python parent id": 3043, "Python module id": 36, "Python id": 3045, "finished": true, "Ev Idx": 9875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564844012.453, "dur": 125.320,
    "args": {
      "Python parent id": 3045, "Python id": 3046, "finished": true, "Ev Idx": 9876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564844012.592, "dur": 0.160,
    "args": {
      "Python parent id": 3046, "Python id": 3047, "finished": true, "Ev Idx": 9877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564844013.476, "dur": 122.903,
    "args": {
      "Python parent id": 3046, "Python id": 3048, "finished": true, "Ev Idx": 9878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564844014.022, "dur": 1.739,
    "args": {
      "Python parent id": 3048, "Python id": 3049, "finished": true, "Ev Idx": 9879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564844016.007, "dur": 11.183,
    "args": {
      "Python parent id": 3048, "Python id": 3050, "finished": true, "Ev Idx": 9880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564844027.668, "dur": 14.605,
    "args": {
      "Python parent id": 3048, "Python id": 3051, "finished": true, "Ev Idx": 9881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564844054.629, "dur": 10.468,
    "args": {
      "Python parent id": 3048, "Python id": 3052, "finished": true, "Ev Idx": 9882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844076.559, "dur": 0.285,
    "args": {
      "Python parent id": 3048, "Python id": 3053, "finished": true, "Ev Idx": 9883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564844077.296, "dur": 1.089,
    "args": {
      "Python parent id": 3048, "Python id": 3054, "finished": true, "Ev Idx": 9884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844139.170, "dur": 0.516,
    "args": {
      "Python parent id": 3043, "Python id": 3055, "finished": true, "Ev Idx": 9885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_18", "pid": 537, "tid": 537,
    "ts": 7429564844143.177, "dur": 668.366,
    "args": {
      "Python parent id": 3043, "Python module id": 18, "Python id": 3056, "finished": true, "Ev Idx": 9886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564844144.868, "dur": 666.109,
    "args": {
      "Python parent id": 3056, "Python id": 3057, "finished": true, "Ev Idx": 9887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564844145.100, "dur": 0.500,
    "args": {
      "Python parent id": 3057, "Python id": 3058, "finished": true, "Ev Idx": 9888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564844147.359, "dur": 662.895,
    "args": {
      "Python parent id": 3057, "Python id": 3059, "finished": true, "Ev Idx": 9889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844151.324, "dur": 0.347,
    "args": {
      "Python parent id": 3059, "Python id": 3060, "finished": true, "Ev Idx": 9890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_126", "pid": 537, "tid": 537,
    "ts": 7429564844154.021, "dur": 52.762,
    "args": {
      "Python parent id": 3059, "Python module id": 126, "Python id": 3061, "finished": true, "Ev Idx": 9891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564844154.680, "dur": 51.868,
    "args": {
      "Python parent id": 3061, "Python id": 3062, "finished": true, "Ev Idx": 9892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564844154.828, "dur": 0.161,
    "args": {
      "Python parent id": 3062, "Python id": 3063, "finished": true, "Ev Idx": 9893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564844155.791, "dur": 50.575,
    "args": {
      "Python parent id": 3062, "Python id": 3064, "finished": true, "Ev Idx": 9894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844156.097, "dur": 0.116,
    "args": {
      "Python parent id": 3064, "Python id": 3065, "finished": true, "Ev Idx": 9895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844156.426, "dur": 0.153,
    "args": {
      "Python parent id": 3064, "Python id": 3066, "finished": true, "Ev Idx": 9896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564844156.730, "dur": 49.516,
    "args": {
      "Python parent id": 3064, "Python id": 3067, "finished": true, "Ev Idx": 9897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564844207.618, "dur": 3.813,
    "args": {
      "Python parent id": 3059, "Python id": 3068, "finished": true, "Ev Idx": 9898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564844211.973, "dur": 3.961,
    "args": {
      "Python parent id": 3059, "Python id": 3069, "finished": true, "Ev Idx": 9899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844216.702, "dur": 0.362,
    "args": {
      "Python parent id": 3059, "Python id": 3070, "finished": true, "Ev Idx": 9900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_127", "pid": 537, "tid": 537,
    "ts": 7429564844219.150, "dur": 47.047,
    "args": {
      "Python parent id": 3059, "Python module id": 127, "Python id": 3071, "finished": true, "Ev Idx": 9901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564844219.709, "dur": 46.269,
    "args": {
      "Python parent id": 3071, "Python id": 3072, "finished": true, "Ev Idx": 9902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564844219.875, "dur": 0.339,
    "args": {
      "Python parent id": 3072, "Python id": 3073, "finished": true, "Ev Idx": 9903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564844221.066, "dur": 44.821,
    "args": {
      "Python parent id": 3072, "Python id": 3074, "finished": true, "Ev Idx": 9904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844221.304, "dur": 0.124,
    "args": {
      "Python parent id": 3074, "Python id": 3075, "finished": true, "Ev Idx": 9905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844221.605, "dur": 0.120,
    "args": {
      "Python parent id": 3074, "Python id": 3076, "finished": true, "Ev Idx": 9906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564844221.831, "dur": 44.002,
    "args": {
      "Python parent id": 3074, "Python id": 3077, "finished": true, "Ev Idx": 9907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564844266.734, "dur": 2.449,
    "args": {
      "Python parent id": 3059, "Python id": 3078, "finished": true, "Ev Idx": 9908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564844269.639, "dur": 2.569,
    "args": {
      "Python parent id": 3059, "Python id": 3079, "finished": true, "Ev Idx": 9909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844272.961, "dur": 0.460,
    "args": {
      "Python parent id": 3059, "Python id": 3080, "finished": true, "Ev Idx": 9910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_128", "pid": 537, "tid": 537,
    "ts": 7429564844275.617, "dur": 43.509,
    "args": {
      "Python parent id": 3059, "Python module id": 128, "Python id": 3081, "finished": true, "Ev Idx": 9911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564844276.157, "dur": 42.758,
    "args": {
      "Python parent id": 3081, "Python id": 3082, "finished": true, "Ev Idx": 9912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564844276.367, "dur": 0.221,
    "args": {
      "Python parent id": 3082, "Python id": 3083, "finished": true, "Ev Idx": 9913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564844277.231, "dur": 41.590,
    "args": {
      "Python parent id": 3082, "Python id": 3084, "finished": true, "Ev Idx": 9914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844278.955, "dur": 0.184,
    "args": {
      "Python parent id": 3084, "Python id": 3085, "finished": true, "Ev Idx": 9915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844279.307, "dur": 0.135,
    "args": {
      "Python parent id": 3084, "Python id": 3086, "finished": true, "Ev Idx": 9916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564844279.544, "dur": 39.182,
    "args": {
      "Python parent id": 3084, "Python id": 3087, "finished": true, "Ev Idx": 9917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564844319.821, "dur": 2.627,
    "args": {
      "Python parent id": 3059, "Python id": 3088, "finished": true, "Ev Idx": 9918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564844322.910, "dur": 2.711,
    "args": {
      "Python parent id": 3059, "Python id": 3089, "finished": true, "Ev Idx": 9919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564844326.809, "dur": 154.217,
    "args": {
      "Python parent id": 3059, "Python id": 3090, "finished": true, "Ev Idx": 9920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564844327.139, "dur": 3.603,
    "args": {
      "Python parent id": 3090, "Python id": 3091, "finished": true, "Ev Idx": 9921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564844331.025, "dur": 2.035,
    "args": {
      "Python parent id": 3090, "Python id": 3092, "finished": true, "Ev Idx": 9922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564844344.619, "dur": 51.744,
    "args": {
      "Python parent id": 3090, "Python id": 3093, "finished": true, "Ev Idx": 9923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564844374.544, "dur": 20.487,
    "args": {
      "Python parent id": 3093, "Python id": 3094, "finished": true, "Ev Idx": 9924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564844430.676, "dur": 32.597,
    "args": {
      "Python parent id": 3090, "Python id": 3095, "finished": true, "Ev Idx": 9925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564844447.235, "dur": 15.063,
    "args": {
      "Python parent id": 3095, "Python id": 3096, "finished": true, "Ev Idx": 9926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564844485.394, "dur": 3.851,
    "args": {
      "Python parent id": 3059, "Python id": 3097, "finished": true, "Ev Idx": 9927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564844486.026, "dur": 0.374,
    "args": {
      "Python parent id": 3097, "Python id": 3098, "finished": true, "Ev Idx": 9928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564844486.697, "dur": 0.088,
    "args": {
      "Python parent id": 3097, "Python id": 3099, "finished": true, "Ev Idx": 9929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564844488.100, "dur": 0.289,
    "args": {
      "Python parent id": 3097, "Python id": 3100, "finished": true, "Ev Idx": 9930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564844488.808, "dur": 0.144,
    "args": {
      "Python parent id": 3097, "Python id": 3101, "finished": true, "Ev Idx": 9931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564844490.662, "dur": 6.778,
    "args": {
      "Python parent id": 3059, "Python id": 3102, "finished": true, "Ev Idx": 9932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564844493.812, "dur": 3.479,
    "args": {
      "Python parent id": 3102, "Python id": 3103, "finished": true, "Ev Idx": 9933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564844494.064, "dur": 1.641,
    "args": {
      "Python parent id": 3103, "Python id": 3104, "finished": true, "Ev Idx": 9934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564844494.761, "dur": 0.774,
    "args": {
      "Python parent id": 3104, "Python id": 3105, "finished": true, "Ev Idx": 9935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564844495.978, "dur": 0.516,
    "args": {
      "Python parent id": 3103, "Python id": 3106, "finished": true, "Ev Idx": 9936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564844496.793, "dur": 0.424,
    "args": {
      "Python parent id": 3103, "Python id": 3107, "finished": true, "Ev Idx": 9937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564844500.107, "dur": 257.230,
    "args": {
      "Python parent id": 3059, "Python id": 3108, "finished": true, "Ev Idx": 9938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564844500.708, "dur": 48.704,
    "args": {
      "Python parent id": 3108, "Python id": 3109, "finished": true, "Ev Idx": 9939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564844519.621, "dur": 4.956,
    "args": {
      "Python parent id": 3109, "Python id": 3110, "finished": true, "Ev Idx": 9940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564844525.478, "dur": 23.804,
    "args": {
      "Python parent id": 3109, "Python id": 3111, "finished": true, "Ev Idx": 9941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564844550.221, "dur": 33.847,
    "args": {
      "Python parent id": 3108, "Python id": 3112, "finished": true, "Ev Idx": 9942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564844567.240, "dur": 3.475,
    "args": {
      "Python parent id": 3112, "Python id": 3113, "finished": true, "Ev Idx": 9943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564844571.195, "dur": 12.809,
    "args": {
      "Python parent id": 3112, "Python id": 3114, "finished": true, "Ev Idx": 9944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564844585.112, "dur": 2.893,
    "args": {
      "Python parent id": 3108, "Python id": 3115, "finished": true, "Ev Idx": 9945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564844588.225, "dur": 46.097,
    "args": {
      "Python parent id": 3108, "Python id": 3116, "finished": true, "Ev Idx": 9946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564844667.660, "dur": 16.657,
    "args": {
      "Python parent id": 3108, "Python id": 3117, "finished": true, "Ev Idx": 9947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564844667.820, "dur": 0.259,
    "args": {
      "Python parent id": 3117, "Python id": 3118, "finished": true, "Ev Idx": 9948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564844668.721, "dur": 15.451,
    "args": {
      "Python parent id": 3117, "Python id": 3119, "finished": true, "Ev Idx": 9949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564844685.296, "dur": 1.840,
    "args": {
      "Python parent id": 3108, "Python id": 3120, "finished": true, "Ev Idx": 9950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564844688.566, "dur": 9.287,
    "args": {
      "Python parent id": 3108, "Python id": 3121, "finished": true, "Ev Idx": 9951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564844688.751, "dur": 0.053,
    "args": {
      "Python parent id": 3121, "Python id": 3122, "finished": true, "Ev Idx": 9952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844693.742, "dur": 0.933,
    "args": {
      "Python parent id": 3121, "Python id": 3123, "finished": true, "Ev Idx": 9953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564844694.165, "dur": 0.432,
    "args": {
      "Python parent id": 3123, "Python id": 3124, "finished": true, "Ev Idx": 9954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564844695.086, "dur": 2.698,
    "args": {
      "Python parent id": 3121, "Python id": 3125, "finished": true, "Ev Idx": 9955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564844698.543, "dur": 39.860,
    "args": {
      "Python parent id": 3108, "Python id": 3126, "finished": true, "Ev Idx": 9956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564844738.946, "dur": 2.926,
    "args": {
      "Python parent id": 3108, "Python id": 3127, "finished": true, "Ev Idx": 9957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564844742.663, "dur": 13.598,
    "args": {
      "Python parent id": 3108, "Python id": 3128, "finished": true, "Ev Idx": 9958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564844760.093, "dur": 3.269,
    "args": {
      "Python parent id": 3059, "Python id": 3129, "finished": true, "Ev Idx": 9959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564844764.065, "dur": 0.258,
    "args": {
      "Python parent id": 3059, "Python id": 3130, "finished": true, "Ev Idx": 9960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844765.304, "dur": 0.743,
    "args": {
      "Python parent id": 3059, "Python id": 3131, "finished": true, "Ev Idx": 9961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_129", "pid": 537, "tid": 537,
    "ts": 7429564844770.116, "dur": 39.589,
    "args": {
      "Python parent id": 3059, "Python module id": 129, "Python id": 3132, "finished": true, "Ev Idx": 9962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564844770.902, "dur": 38.477,
    "args": {
      "Python parent id": 3132, "Python id": 3133, "finished": true, "Ev Idx": 9963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564844771.258, "dur": 0.922,
    "args": {
      "Python parent id": 3133, "Python id": 3134, "finished": true, "Ev Idx": 9964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564844773.243, "dur": 35.858,
    "args": {
      "Python parent id": 3133, "Python id": 3135, "finished": true, "Ev Idx": 9965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844773.617, "dur": 0.211,
    "args": {
      "Python parent id": 3135, "Python id": 3136, "finished": true, "Ev Idx": 9966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844774.051, "dur": 0.151,
    "args": {
      "Python parent id": 3135, "Python id": 3137, "finished": true, "Ev Idx": 9967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564844774.368, "dur": 34.624,
    "args": {
      "Python parent id": 3135, "Python id": 3138, "finished": true, "Ev Idx": 9968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844824.344, "dur": 0.606,
    "args": {
      "Python parent id": 3043, "Python id": 3139, "finished": true, "Ev Idx": 9969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_37", "pid": 537, "tid": 537,
    "ts": 7429564844827.827, "dur": 83.806,
    "args": {
      "Python parent id": 3043, "Python module id": 37, "Python id": 3140, "finished": true, "Ev Idx": 9970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564844828.444, "dur": 82.736,
    "args": {
      "Python parent id": 3140, "Python id": 3141, "finished": true, "Ev Idx": 9971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564844828.648, "dur": 0.431,
    "args": {
      "Python parent id": 3141, "Python id": 3142, "finished": true, "Ev Idx": 9972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564844830.231, "dur": 80.091,
    "args": {
      "Python parent id": 3141, "Python id": 3143, "finished": true, "Ev Idx": 9973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564844830.706, "dur": 1.755,
    "args": {
      "Python parent id": 3143, "Python id": 3144, "finished": true, "Ev Idx": 9974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564844832.781, "dur": 14.040,
    "args": {
      "Python parent id": 3143, "Python id": 3145, "finished": true, "Ev Idx": 9975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564844847.165, "dur": 15.479,
    "args": {
      "Python parent id": 3143, "Python id": 3146, "finished": true, "Ev Idx": 9976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564844875.522, "dur": 12.039,
    "args": {
      "Python parent id": 3143, "Python id": 3147, "finished": true, "Ev Idx": 9977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844897.625, "dur": 0.231,
    "args": {
      "Python parent id": 3143, "Python id": 3148, "finished": true, "Ev Idx": 9978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564844898.228, "dur": 1.197,
    "args": {
      "Python parent id": 3143, "Python id": 3149, "finished": true, "Ev Idx": 9979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844912.362, "dur": 0.531,
    "args": {
      "Python parent id": 3043, "Python id": 3150, "finished": true, "Ev Idx": 9980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_18", "pid": 537, "tid": 537,
    "ts": 7429564844914.508, "dur": 282.683,
    "args": {
      "Python parent id": 3043, "Python module id": 18, "Python id": 3151, "finished": true, "Ev Idx": 9981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564844915.068, "dur": 281.814,
    "args": {
      "Python parent id": 3151, "Python id": 3152, "finished": true, "Ev Idx": 9982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564844915.242, "dur": 0.330,
    "args": {
      "Python parent id": 3152, "Python id": 3153, "finished": true, "Ev Idx": 9983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564844916.572, "dur": 280.180,
    "args": {
      "Python parent id": 3152, "Python id": 3154, "finished": true, "Ev Idx": 9984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844916.856, "dur": 0.311,
    "args": {
      "Python parent id": 3154, "Python id": 3155, "finished": true, "Ev Idx": 9985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844917.502, "dur": 0.297,
    "args": {
      "Python parent id": 3154, "Python id": 3156, "finished": true, "Ev Idx": 9986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844918.045, "dur": 0.208,
    "args": {
      "Python parent id": 3154, "Python id": 3157, "finished": true, "Ev Idx": 9987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_130", "pid": 537, "tid": 537,
    "ts": 7429564844919.887, "dur": 111.173,
    "args": {
      "Python parent id": 3154, "Python module id": 130, "Python id": 3158, "finished": true, "Ev Idx": 9988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564844920.542, "dur": 110.231,
    "args": {
      "Python parent id": 3158, "Python id": 3159, "finished": true, "Ev Idx": 9989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564844920.688, "dur": 0.178,
    "args": {
      "Python parent id": 3159, "Python id": 3160, "finished": true, "Ev Idx": 9990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564844921.741, "dur": 108.828,
    "args": {
      "Python parent id": 3159, "Python id": 3161, "finished": true, "Ev Idx": 9991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844922.034, "dur": 0.124,
    "args": {
      "Python parent id": 3161, "Python id": 3162, "finished": true, "Ev Idx": 9992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564844922.376, "dur": 0.111,
    "args": {
      "Python parent id": 3161, "Python id": 3163, "finished": true, "Ev Idx": 9993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564844922.616, "dur": 107.858,
    "args": {
      "Python parent id": 3161, "Python id": 3164, "finished": true, "Ev Idx": 9994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_18", "pid": 537, "tid": 537,
    "ts": 7429564845033.311, "dur": 18.457,
    "args": {
      "Python parent id": 3154, "Python module id": 18, "Python id": 3165, "finished": true, "Ev Idx": 9995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564845033.861, "dur": 17.614,
    "args": {
      "Python parent id": 3165, "Python id": 3166, "finished": true, "Ev Idx": 9996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564845034.059, "dur": 0.459,
    "args": {
      "Python parent id": 3166, "Python id": 3167, "finished": true, "Ev Idx": 9997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564845035.775, "dur": 15.567,
    "args": {
      "Python parent id": 3166, "Python id": 3168, "finished": true, "Ev Idx": 9998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564845036.830, "dur": 14.339,
    "args": {
      "Python parent id": 3168, "Python id": 3169, "finished": true, "Ev Idx": 9999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564845036.976, "dur": 0.288,
    "args": {
      "Python parent id": 3169, "Python id": 3170, "finished": true, "Ev Idx": 10000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564845038.224, "dur": 12.845,
    "args": {
      "Python parent id": 3169, "Python id": 3171, "finished": true, "Ev Idx": 10001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845052.518, "dur": 0.465,
    "args": {
      "Python parent id": 3154, "Python id": 3172, "finished": true, "Ev Idx": 10002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_131", "pid": 537, "tid": 537,
    "ts": 7429564845054.998, "dur": 82.893,
    "args": {
      "Python parent id": 3154, "Python module id": 131, "Python id": 3173, "finished": true, "Ev Idx": 10003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564845055.420, "dur": 82.218,
    "args": {
      "Python parent id": 3173, "Python id": 3174, "finished": true, "Ev Idx": 10004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564845055.577, "dur": 0.241,
    "args": {
      "Python parent id": 3174, "Python id": 3175, "finished": true, "Ev Idx": 10005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564845056.776, "dur": 80.678,
    "args": {
      "Python parent id": 3174, "Python id": 3176, "finished": true, "Ev Idx": 10006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845057.030, "dur": 0.163,
    "args": {
      "Python parent id": 3176, "Python id": 3177, "finished": true, "Ev Idx": 10007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845057.443, "dur": 0.125,
    "args": {
      "Python parent id": 3176, "Python id": 3178, "finished": true, "Ev Idx": 10008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564845057.699, "dur": 79.675,
    "args": {
      "Python parent id": 3176, "Python id": 3179, "finished": true, "Ev Idx": 10009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_132", "pid": 537, "tid": 537,
    "ts": 7429564845152.207, "dur": 44.230,
    "args": {
      "Python parent id": 3154, "Python module id": 132, "Python id": 3180, "finished": true, "Ev Idx": 10010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564845152.986, "dur": 43.260,
    "args": {
      "Python parent id": 3180, "Python id": 3181, "finished": true, "Ev Idx": 10011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564845153.307, "dur": 0.497,
    "args": {
      "Python parent id": 3181, "Python id": 3182, "finished": true, "Ev Idx": 10012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564845154.728, "dur": 41.409,
    "args": {
      "Python parent id": 3181, "Python id": 3183, "finished": true, "Ev Idx": 10013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845155.015, "dur": 0.186,
    "args": {
      "Python parent id": 3183, "Python id": 3184, "finished": true, "Ev Idx": 10014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845155.426, "dur": 0.135,
    "args": {
      "Python parent id": 3183, "Python id": 3185, "finished": true, "Ev Idx": 10015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564845155.713, "dur": 40.361,
    "args": {
      "Python parent id": 3183, "Python id": 3186, "finished": true, "Ev Idx": 10016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_19", "pid": 537, "tid": 537,
    "ts": 7429564845218.231, "dur": 1277.868,
    "args": {
      "Python parent id": 64, "Python module id": 19, "Python id": 3187, "finished": true, "Ev Idx": 10017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564845219.733, "dur": 1275.670,
    "args": {
      "Python parent id": 3187, "Python id": 3188, "finished": true, "Ev Idx": 10018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564845220.014, "dur": 0.389,
    "args": {
      "Python parent id": 3188, "Python id": 3189, "finished": true, "Ev Idx": 10019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564845222.114, "dur": 1272.431,
    "args": {
      "Python parent id": 3188, "Python id": 3190, "finished": true, "Ev Idx": 10020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845222.641, "dur": 0.742,
    "args": {
      "Python parent id": 3190, "Python id": 3191, "finished": true, "Ev Idx": 10021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_38", "pid": 537, "tid": 537,
    "ts": 7429564845225.814, "dur": 112.854,
    "args": {
      "Python parent id": 3190, "Python module id": 38, "Python id": 3192, "finished": true, "Ev Idx": 10022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564845226.454, "dur": 111.862,
    "args": {
      "Python parent id": 3192, "Python id": 3193, "finished": true, "Ev Idx": 10023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564845226.650, "dur": 0.166,
    "args": {
      "Python parent id": 3193, "Python id": 3194, "finished": true, "Ev Idx": 10024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564845227.756, "dur": 109.865,
    "args": {
      "Python parent id": 3193, "Python id": 3195, "finished": true, "Ev Idx": 10025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564845228.311, "dur": 2.059,
    "args": {
      "Python parent id": 3195, "Python id": 3196, "finished": true, "Ev Idx": 10026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564845230.675, "dur": 13.214,
    "args": {
      "Python parent id": 3195, "Python id": 3197, "finished": true, "Ev Idx": 10027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564845244.351, "dur": 14.878,
    "args": {
      "Python parent id": 3195, "Python id": 3198, "finished": true, "Ev Idx": 10028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564845271.796, "dur": 9.932,
    "args": {
      "Python parent id": 3195, "Python id": 3199, "finished": true, "Ev Idx": 10029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845293.887, "dur": 0.194,
    "args": {
      "Python parent id": 3195, "Python id": 3200, "finished": true, "Ev Idx": 10030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564845294.670, "dur": 34.006,
    "args": {
      "Python parent id": 3195, "Python id": 3201, "finished": true, "Ev Idx": 10031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845339.300, "dur": 0.429,
    "args": {
      "Python parent id": 3190, "Python id": 3202, "finished": true, "Ev Idx": 10032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_19", "pid": 537, "tid": 537,
    "ts": 7429564845342.504, "dur": 713.130,
    "args": {
      "Python parent id": 3190, "Python module id": 19, "Python id": 3203, "finished": true, "Ev Idx": 10033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564845344.046, "dur": 710.930,
    "args": {
      "Python parent id": 3203, "Python id": 3204, "finished": true, "Ev Idx": 10034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564845344.245, "dur": 0.229,
    "args": {
      "Python parent id": 3204, "Python id": 3205, "finished": true, "Ev Idx": 10035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564845346.180, "dur": 708.023,
    "args": {
      "Python parent id": 3204, "Python id": 3206, "finished": true, "Ev Idx": 10036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845349.743, "dur": 0.313,
    "args": {
      "Python parent id": 3206, "Python id": 3207, "finished": true, "Ev Idx": 10037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_133", "pid": 537, "tid": 537,
    "ts": 7429564845352.315, "dur": 42.503,
    "args": {
      "Python parent id": 3206, "Python module id": 133, "Python id": 3208, "finished": true, "Ev Idx": 10038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564845352.847, "dur": 41.710,
    "args": {
      "Python parent id": 3208, "Python id": 3209, "finished": true, "Ev Idx": 10039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564845353.005, "dur": 0.165,
    "args": {
      "Python parent id": 3209, "Python id": 3210, "finished": true, "Ev Idx": 10040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564845353.957, "dur": 40.443,
    "args": {
      "Python parent id": 3209, "Python id": 3211, "finished": true, "Ev Idx": 10041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845354.299, "dur": 0.205,
    "args": {
      "Python parent id": 3211, "Python id": 3212, "finished": true, "Ev Idx": 10042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845354.722, "dur": 0.130,
    "args": {
      "Python parent id": 3211, "Python id": 3213, "finished": true, "Ev Idx": 10043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564845354.957, "dur": 39.371,
    "args": {
      "Python parent id": 3211, "Python id": 3214, "finished": true, "Ev Idx": 10044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564845395.410, "dur": 3.339,
    "args": {
      "Python parent id": 3206, "Python id": 3215, "finished": true, "Ev Idx": 10045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564845399.278, "dur": 3.995,
    "args": {
      "Python parent id": 3206, "Python id": 3216, "finished": true, "Ev Idx": 10046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845404.090, "dur": 0.391,
    "args": {
      "Python parent id": 3206, "Python id": 3217, "finished": true, "Ev Idx": 10047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_134", "pid": 537, "tid": 537,
    "ts": 7429564845406.601, "dur": 43.934,
    "args": {
      "Python parent id": 3206, "Python module id": 134, "Python id": 3218, "finished": true, "Ev Idx": 10048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564845407.136, "dur": 43.133,
    "args": {
      "Python parent id": 3218, "Python id": 3219, "finished": true, "Ev Idx": 10049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564845407.383, "dur": 0.301,
    "args": {
      "Python parent id": 3219, "Python id": 3220, "finished": true, "Ev Idx": 10050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564845408.386, "dur": 41.746,
    "args": {
      "Python parent id": 3219, "Python id": 3221, "finished": true, "Ev Idx": 10051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845408.627, "dur": 0.396,
    "args": {
      "Python parent id": 3221, "Python id": 3222, "finished": true, "Ev Idx": 10052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845409.203, "dur": 0.136,
    "args": {
      "Python parent id": 3221, "Python id": 3223, "finished": true, "Ev Idx": 10053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564845409.460, "dur": 40.619,
    "args": {
      "Python parent id": 3221, "Python id": 3224, "finished": true, "Ev Idx": 10054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564845451.107, "dur": 2.446,
    "args": {
      "Python parent id": 3206, "Python id": 3225, "finished": true, "Ev Idx": 10055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564845454.025, "dur": 2.858,
    "args": {
      "Python parent id": 3206, "Python id": 3226, "finished": true, "Ev Idx": 10056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845457.613, "dur": 0.427,
    "args": {
      "Python parent id": 3206, "Python id": 3227, "finished": true, "Ev Idx": 10057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_135", "pid": 537, "tid": 537,
    "ts": 7429564845484.902, "dur": 36.577,
    "args": {
      "Python parent id": 3206, "Python module id": 135, "Python id": 3228, "finished": true, "Ev Idx": 10058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564845485.560, "dur": 35.570,
    "args": {
      "Python parent id": 3228, "Python id": 3229, "finished": true, "Ev Idx": 10059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564845485.767, "dur": 0.362,
    "args": {
      "Python parent id": 3229, "Python id": 3230, "finished": true, "Ev Idx": 10060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564845486.825, "dur": 34.182,
    "args": {
      "Python parent id": 3229, "Python id": 3231, "finished": true, "Ev Idx": 10061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845487.039, "dur": 0.134,
    "args": {
      "Python parent id": 3231, "Python id": 3232, "finished": true, "Ev Idx": 10062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845487.379, "dur": 0.125,
    "args": {
      "Python parent id": 3231, "Python id": 3233, "finished": true, "Ev Idx": 10063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564845487.614, "dur": 33.314,
    "args": {
      "Python parent id": 3231, "Python id": 3234, "finished": true, "Ev Idx": 10064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564845522.105, "dur": 2.547,
    "args": {
      "Python parent id": 3206, "Python id": 3235, "finished": true, "Ev Idx": 10065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564845525.269, "dur": 2.492,
    "args": {
      "Python parent id": 3206, "Python id": 3236, "finished": true, "Ev Idx": 10066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564845529.062, "dur": 188.640,
    "args": {
      "Python parent id": 3206, "Python id": 3237, "finished": true, "Ev Idx": 10067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564845529.356, "dur": 3.632,
    "args": {
      "Python parent id": 3237, "Python id": 3238, "finished": true, "Ev Idx": 10068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564845533.316, "dur": 1.756,
    "args": {
      "Python parent id": 3237, "Python id": 3239, "finished": true, "Ev Idx": 10069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564845546.836, "dur": 92.742,
    "args": {
      "Python parent id": 3237, "Python id": 3240, "finished": true, "Ev Idx": 10070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564845573.532, "dur": 64.695,
    "args": {
      "Python parent id": 3240, "Python id": 3241, "finished": true, "Ev Idx": 10071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564845670.158, "dur": 30.382,
    "args": {
      "Python parent id": 3237, "Python id": 3242, "finished": true, "Ev Idx": 10072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564845688.072, "dur": 11.858,
    "args": {
      "Python parent id": 3242, "Python id": 3243, "finished": true, "Ev Idx": 10073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564845722.335, "dur": 4.119,
    "args": {
      "Python parent id": 3206, "Python id": 3244, "finished": true, "Ev Idx": 10074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564845723.117, "dur": 0.347,
    "args": {
      "Python parent id": 3244, "Python id": 3245, "finished": true, "Ev Idx": 10075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564845723.913, "dur": 0.039,
    "args": {
      "Python parent id": 3244, "Python id": 3246, "finished": true, "Ev Idx": 10076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564845725.430, "dur": 0.183,
    "args": {
      "Python parent id": 3244, "Python id": 3247, "finished": true, "Ev Idx": 10077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564845726.055, "dur": 0.139,
    "args": {
      "Python parent id": 3244, "Python id": 3248, "finished": true, "Ev Idx": 10078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564845727.913, "dur": 7.077,
    "args": {
      "Python parent id": 3206, "Python id": 3249, "finished": true, "Ev Idx": 10079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564845731.395, "dur": 3.442,
    "args": {
      "Python parent id": 3249, "Python id": 3250, "finished": true, "Ev Idx": 10080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564845731.626, "dur": 1.564,
    "args": {
      "Python parent id": 3250, "Python id": 3251, "finished": true, "Ev Idx": 10081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564845732.180, "dur": 0.819,
    "args": {
      "Python parent id": 3251, "Python id": 3252, "finished": true, "Ev Idx": 10082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564845733.482, "dur": 0.509,
    "args": {
      "Python parent id": 3250, "Python id": 3253, "finished": true, "Ev Idx": 10083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564845734.313, "dur": 0.452,
    "args": {
      "Python parent id": 3250, "Python id": 3254, "finished": true, "Ev Idx": 10084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564845738.067, "dur": 260.716,
    "args": {
      "Python parent id": 3206, "Python id": 3255, "finished": true, "Ev Idx": 10085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564845738.625, "dur": 46.584,
    "args": {
      "Python parent id": 3255, "Python id": 3256, "finished": true, "Ev Idx": 10086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564845755.768, "dur": 5.418,
    "args": {
      "Python parent id": 3256, "Python id": 3257, "finished": true, "Ev Idx": 10087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564845762.133, "dur": 22.943,
    "args": {
      "Python parent id": 3256, "Python id": 3258, "finished": true, "Ev Idx": 10088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564845786.150, "dur": 28.820,
    "args": {
      "Python parent id": 3255, "Python id": 3259, "finished": true, "Ev Idx": 10089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564845799.413, "dur": 2.749,
    "args": {
      "Python parent id": 3259, "Python id": 3260, "finished": true, "Ev Idx": 10090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564845802.674, "dur": 12.203,
    "args": {
      "Python parent id": 3259, "Python id": 3261, "finished": true, "Ev Idx": 10091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564845815.974, "dur": 3.165,
    "args": {
      "Python parent id": 3255, "Python id": 3262, "finished": true, "Ev Idx": 10092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564845819.423, "dur": 50.387,
    "args": {
      "Python parent id": 3255, "Python id": 3263, "finished": true, "Ev Idx": 10093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564845904.337, "dur": 17.972,
    "args": {
      "Python parent id": 3255, "Python id": 3264, "finished": true, "Ev Idx": 10094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564845904.646, "dur": 0.235,
    "args": {
      "Python parent id": 3264, "Python id": 3265, "finished": true, "Ev Idx": 10095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564845905.614, "dur": 16.543,
    "args": {
      "Python parent id": 3264, "Python id": 3266, "finished": true, "Ev Idx": 10096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564845923.236, "dur": 1.859,
    "args": {
      "Python parent id": 3255, "Python id": 3267, "finished": true, "Ev Idx": 10097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564845926.656, "dur": 9.835,
    "args": {
      "Python parent id": 3255, "Python id": 3268, "finished": true, "Ev Idx": 10098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564845926.833, "dur": 0.057,
    "args": {
      "Python parent id": 3268, "Python id": 3269, "finished": true, "Ev Idx": 10099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564845932.367, "dur": 0.914,
    "args": {
      "Python parent id": 3268, "Python id": 3270, "finished": true, "Ev Idx": 10100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564845932.820, "dur": 0.366,
    "args": {
      "Python parent id": 3270, "Python id": 3271, "finished": true, "Ev Idx": 10101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564845933.739, "dur": 2.667,
    "args": {
      "Python parent id": 3268, "Python id": 3272, "finished": true, "Ev Idx": 10102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564845937.232, "dur": 42.674,
    "args": {
      "Python parent id": 3255, "Python id": 3273, "finished": true, "Ev Idx": 10103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564845980.553, "dur": 3.077,
    "args": {
      "Python parent id": 3255, "Python id": 3274, "finished": true, "Ev Idx": 10104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564845984.341, "dur": 13.441,
    "args": {
      "Python parent id": 3255, "Python id": 3275, "finished": true, "Ev Idx": 10105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564846001.742, "dur": 3.392,
    "args": {
      "Python parent id": 3206, "Python id": 3276, "finished": true, "Ev Idx": 10106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564846005.863, "dur": 0.314,
    "args": {
      "Python parent id": 3206, "Python id": 3277, "finished": true, "Ev Idx": 10107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846007.146, "dur": 0.902,
    "args": {
      "Python parent id": 3206, "Python id": 3278, "finished": true, "Ev Idx": 10108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_136", "pid": 537, "tid": 537,
    "ts": 7429564846012.505, "dur": 41.099,
    "args": {
      "Python parent id": 3206, "Python module id": 136, "Python id": 3279, "finished": true, "Ev Idx": 10109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846013.641, "dur": 39.587,
    "args": {
      "Python parent id": 3279, "Python id": 3280, "finished": true, "Ev Idx": 10110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846014.037, "dur": 1.003,
    "args": {
      "Python parent id": 3280, "Python id": 3281, "finished": true, "Ev Idx": 10111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564846016.258, "dur": 36.714,
    "args": {
      "Python parent id": 3280, "Python id": 3282, "finished": true, "Ev Idx": 10112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846016.630, "dur": 0.237,
    "args": {
      "Python parent id": 3282, "Python id": 3283, "finished": true, "Ev Idx": 10113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846017.102, "dur": 0.169,
    "args": {
      "Python parent id": 3282, "Python id": 3284, "finished": true, "Ev Idx": 10114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564846017.401, "dur": 35.430,
    "args": {
      "Python parent id": 3282, "Python id": 3285, "finished": true, "Ev Idx": 10115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846068.497, "dur": 0.746,
    "args": {
      "Python parent id": 3190, "Python id": 3286, "finished": true, "Ev Idx": 10116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_39", "pid": 537, "tid": 537,
    "ts": 7429564846090.372, "dur": 96.404,
    "args": {
      "Python parent id": 3190, "Python module id": 39, "Python id": 3287, "finished": true, "Ev Idx": 10117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846091.169, "dur": 95.193,
    "args": {
      "Python parent id": 3287, "Python id": 3288, "finished": true, "Ev Idx": 10118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846091.389, "dur": 0.672,
    "args": {
      "Python parent id": 3288, "Python id": 3289, "finished": true, "Ev Idx": 10119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564846093.241, "dur": 92.243,
    "args": {
      "Python parent id": 3288, "Python id": 3290, "finished": true, "Ev Idx": 10120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564846093.908, "dur": 3.179,
    "args": {
      "Python parent id": 3290, "Python id": 3291, "finished": true, "Ev Idx": 10121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564846097.402, "dur": 18.500,
    "args": {
      "Python parent id": 3290, "Python id": 3292, "finished": true, "Ev Idx": 10122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564846116.380, "dur": 17.053,
    "args": {
      "Python parent id": 3290, "Python id": 3293, "finished": true, "Ev Idx": 10123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564846148.612, "dur": 11.979,
    "args": {
      "Python parent id": 3290, "Python id": 3294, "finished": true, "Ev Idx": 10124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846172.356, "dur": 0.279,
    "args": {
      "Python parent id": 3290, "Python id": 3295, "finished": true, "Ev Idx": 10125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564846173.039, "dur": 1.296,
    "args": {
      "Python parent id": 3290, "Python id": 3296, "finished": true, "Ev Idx": 10126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846187.515, "dur": 0.401,
    "args": {
      "Python parent id": 3190, "Python id": 3297, "finished": true, "Ev Idx": 10127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_19", "pid": 537, "tid": 537,
    "ts": 7429564846189.898, "dur": 292.773,
    "args": {
      "Python parent id": 3190, "Python module id": 19, "Python id": 3298, "finished": true, "Ev Idx": 10128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846190.796, "dur": 291.627,
    "args": {
      "Python parent id": 3298, "Python id": 3299, "finished": true, "Ev Idx": 10129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846190.983, "dur": 0.409,
    "args": {
      "Python parent id": 3299, "Python id": 3300, "finished": true, "Ev Idx": 10130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564846192.531, "dur": 289.765,
    "args": {
      "Python parent id": 3299, "Python id": 3301, "finished": true, "Ev Idx": 10131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846192.860, "dur": 0.355,
    "args": {
      "Python parent id": 3301, "Python id": 3302, "finished": true, "Ev Idx": 10132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846193.611, "dur": 0.283,
    "args": {
      "Python parent id": 3301, "Python id": 3303, "finished": true, "Ev Idx": 10133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846194.188, "dur": 0.337,
    "args": {
      "Python parent id": 3301, "Python id": 3304, "finished": true, "Ev Idx": 10134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_137", "pid": 537, "tid": 537,
    "ts": 7429564846196.594, "dur": 59.200,
    "args": {
      "Python parent id": 3301, "Python module id": 137, "Python id": 3305, "finished": true, "Ev Idx": 10135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846197.105, "dur": 58.393,
    "args": {
      "Python parent id": 3305, "Python id": 3306, "finished": true, "Ev Idx": 10136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846197.241, "dur": 0.211,
    "args": {
      "Python parent id": 3306, "Python id": 3307, "finished": true, "Ev Idx": 10137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564846198.421, "dur": 56.947,
    "args": {
      "Python parent id": 3306, "Python id": 3308, "finished": true, "Ev Idx": 10138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846198.702, "dur": 0.159,
    "args": {
      "Python parent id": 3308, "Python id": 3309, "finished": true, "Ev Idx": 10139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846199.091, "dur": 0.165,
    "args": {
      "Python parent id": 3308, "Python id": 3310, "finished": true, "Ev Idx": 10140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564846199.372, "dur": 55.871,
    "args": {
      "Python parent id": 3308, "Python id": 3311, "finished": true, "Ev Idx": 10141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_19", "pid": 537, "tid": 537,
    "ts": 7429564846257.803, "dur": 76.213,
    "args": {
      "Python parent id": 3301, "Python module id": 19, "Python id": 3312, "finished": true, "Ev Idx": 10142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846258.223, "dur": 75.490,
    "args": {
      "Python parent id": 3312, "Python id": 3313, "finished": true, "Ev Idx": 10143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846258.445, "dur": 0.383,
    "args": {
      "Python parent id": 3313, "Python id": 3314, "finished": true, "Ev Idx": 10144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564846259.824, "dur": 73.723,
    "args": {
      "Python parent id": 3313, "Python id": 3315, "finished": true, "Ev Idx": 10145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564846260.749, "dur": 72.625,
    "args": {
      "Python parent id": 3315, "Python id": 3316, "finished": true, "Ev Idx": 10146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564846260.876, "dur": 0.173,
    "args": {
      "Python parent id": 3316, "Python id": 3317, "finished": true, "Ev Idx": 10147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564846261.977, "dur": 71.277,
    "args": {
      "Python parent id": 3316, "Python id": 3318, "finished": true, "Ev Idx": 10148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846334.979, "dur": 0.417,
    "args": {
      "Python parent id": 3301, "Python id": 3319, "finished": true, "Ev Idx": 10149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_138", "pid": 537, "tid": 537,
    "ts": 7429564846337.228, "dur": 52.272,
    "args": {
      "Python parent id": 3301, "Python module id": 138, "Python id": 3320, "finished": true, "Ev Idx": 10150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846337.685, "dur": 51.630,
    "args": {
      "Python parent id": 3320, "Python id": 3321, "finished": true, "Ev Idx": 10151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846337.873, "dur": 0.359,
    "args": {
      "Python parent id": 3321, "Python id": 3322, "finished": true, "Ev Idx": 10152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564846339.080, "dur": 50.123,
    "args": {
      "Python parent id": 3321, "Python id": 3323, "finished": true, "Ev Idx": 10153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846339.343, "dur": 0.120,
    "args": {
      "Python parent id": 3323, "Python id": 3324, "finished": true, "Ev Idx": 10154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846339.707, "dur": 0.125,
    "args": {
      "Python parent id": 3323, "Python id": 3325, "finished": true, "Ev Idx": 10155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564846339.943, "dur": 49.114,
    "args": {
      "Python parent id": 3323, "Python id": 3326, "finished": true, "Ev Idx": 10156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_139", "pid": 537, "tid": 537,
    "ts": 7429564846433.035, "dur": 48.858,
    "args": {
      "Python parent id": 3301, "Python module id": 139, "Python id": 3327, "finished": true, "Ev Idx": 10157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846433.800, "dur": 47.750,
    "args": {
      "Python parent id": 3327, "Python id": 3328, "finished": true, "Ev Idx": 10158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846434.077, "dur": 0.479,
    "args": {
      "Python parent id": 3328, "Python id": 3329, "finished": true, "Ev Idx": 10159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564846435.295, "dur": 46.116,
    "args": {
      "Python parent id": 3328, "Python id": 3330, "finished": true, "Ev Idx": 10160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846435.637, "dur": 0.196,
    "args": {
      "Python parent id": 3330, "Python id": 3331, "finished": true, "Ev Idx": 10161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846436.058, "dur": 0.124,
    "args": {
      "Python parent id": 3330, "Python id": 3332, "finished": true, "Ev Idx": 10162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564846436.293, "dur": 45.023,
    "args": {
      "Python parent id": 3330, "Python id": 3333, "finished": true, "Ev Idx": 10163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_20", "pid": 537, "tid": 537,
    "ts": 7429564846502.592, "dur": 1171.647,
    "args": {
      "Python parent id": 64, "Python module id": 20, "Python id": 3334, "finished": true, "Ev Idx": 10164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846503.990, "dur": 1169.479,
    "args": {
      "Python parent id": 3334, "Python id": 3335, "finished": true, "Ev Idx": 10165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846504.209, "dur": 0.394,
    "args": {
      "Python parent id": 3335, "Python id": 3336, "finished": true, "Ev Idx": 10166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564846506.264, "dur": 1166.202,
    "args": {
      "Python parent id": 3335, "Python id": 3337, "finished": true, "Ev Idx": 10167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846506.730, "dur": 0.576,
    "args": {
      "Python parent id": 3337, "Python id": 3338, "finished": true, "Ev Idx": 10168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_40", "pid": 537, "tid": 537,
    "ts": 7429564846509.332, "dur": 83.815,
    "args": {
      "Python parent id": 3337, "Python module id": 40, "Python id": 3339, "finished": true, "Ev Idx": 10169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846510.032, "dur": 82.784,
    "args": {
      "Python parent id": 3339, "Python id": 3340, "finished": true, "Ev Idx": 10170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846510.179, "dur": 0.157,
    "args": {
      "Python parent id": 3340, "Python id": 3341, "finished": true, "Ev Idx": 10171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564846511.284, "dur": 80.789,
    "args": {
      "Python parent id": 3340, "Python id": 3342, "finished": true, "Ev Idx": 10172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564846511.898, "dur": 2.292,
    "args": {
      "Python parent id": 3342, "Python id": 3343, "finished": true, "Ev Idx": 10173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564846514.519, "dur": 12.641,
    "args": {
      "Python parent id": 3342, "Python id": 3344, "finished": true, "Ev Idx": 10174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564846527.508, "dur": 15.884,
    "args": {
      "Python parent id": 3342, "Python id": 3345, "finished": true, "Ev Idx": 10175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564846556.980, "dur": 13.094,
    "args": {
      "Python parent id": 3342, "Python id": 3346, "finished": true, "Ev Idx": 10176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846580.985, "dur": 0.223,
    "args": {
      "Python parent id": 3342, "Python id": 3347, "finished": true, "Ev Idx": 10177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564846581.659, "dur": 1.448,
    "args": {
      "Python parent id": 3342, "Python id": 3348, "finished": true, "Ev Idx": 10178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846593.842, "dur": 0.347,
    "args": {
      "Python parent id": 3337, "Python id": 3349, "finished": true, "Ev Idx": 10179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_20", "pid": 537, "tid": 537,
    "ts": 7429564846597.114, "dur": 668.550,
    "args": {
      "Python parent id": 3337, "Python module id": 20, "Python id": 3350, "finished": true, "Ev Idx": 10180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846598.662, "dur": 666.343,
    "args": {
      "Python parent id": 3350, "Python id": 3351, "finished": true, "Ev Idx": 10181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846598.848, "dur": 0.362,
    "args": {
      "Python parent id": 3351, "Python id": 3352, "finished": true, "Ev Idx": 10182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564846600.792, "dur": 663.422,
    "args": {
      "Python parent id": 3351, "Python id": 3353, "finished": true, "Ev Idx": 10183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846604.300, "dur": 0.371,
    "args": {
      "Python parent id": 3353, "Python id": 3354, "finished": true, "Ev Idx": 10184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_140", "pid": 537, "tid": 537,
    "ts": 7429564846606.851, "dur": 45.858,
    "args": {
      "Python parent id": 3353, "Python module id": 140, "Python id": 3355, "finished": true, "Ev Idx": 10185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846607.468, "dur": 44.921,
    "args": {
      "Python parent id": 3355, "Python id": 3356, "finished": true, "Ev Idx": 10186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846607.612, "dur": 0.174,
    "args": {
      "Python parent id": 3356, "Python id": 3357, "finished": true, "Ev Idx": 10187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564846608.514, "dur": 43.754,
    "args": {
      "Python parent id": 3356, "Python id": 3358, "finished": true, "Ev Idx": 10188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846608.913, "dur": 0.213,
    "args": {
      "Python parent id": 3358, "Python id": 3359, "finished": true, "Ev Idx": 10189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846609.382, "dur": 0.154,
    "args": {
      "Python parent id": 3358, "Python id": 3360, "finished": true, "Ev Idx": 10190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564846609.708, "dur": 42.468,
    "args": {
      "Python parent id": 3358, "Python id": 3361, "finished": true, "Ev Idx": 10191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564846653.486, "dur": 3.528,
    "args": {
      "Python parent id": 3353, "Python id": 3362, "finished": true, "Ev Idx": 10192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564846657.440, "dur": 3.921,
    "args": {
      "Python parent id": 3353, "Python id": 3363, "finished": true, "Ev Idx": 10193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846662.344, "dur": 0.422,
    "args": {
      "Python parent id": 3353, "Python id": 3364, "finished": true, "Ev Idx": 10194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_141", "pid": 537, "tid": 537,
    "ts": 7429564846665.128, "dur": 48.995,
    "args": {
      "Python parent id": 3353, "Python module id": 141, "Python id": 3365, "finished": true, "Ev Idx": 10195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846665.606, "dur": 48.286,
    "args": {
      "Python parent id": 3365, "Python id": 3366, "finished": true, "Ev Idx": 10196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846665.811, "dur": 0.292,
    "args": {
      "Python parent id": 3366, "Python id": 3367, "finished": true, "Ev Idx": 10197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564846666.887, "dur": 46.874,
    "args": {
      "Python parent id": 3366, "Python id": 3368, "finished": true, "Ev Idx": 10198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846667.122, "dur": 0.164,
    "args": {
      "Python parent id": 3368, "Python id": 3369, "finished": true, "Ev Idx": 10199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846667.502, "dur": 0.143,
    "args": {
      "Python parent id": 3368, "Python id": 3370, "finished": true, "Ev Idx": 10200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564846667.767, "dur": 45.921,
    "args": {
      "Python parent id": 3368, "Python id": 3371, "finished": true, "Ev Idx": 10201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564846714.766, "dur": 2.445,
    "args": {
      "Python parent id": 3353, "Python id": 3372, "finished": true, "Ev Idx": 10202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564846717.735, "dur": 2.855,
    "args": {
      "Python parent id": 3353, "Python id": 3373, "finished": true, "Ev Idx": 10203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846721.371, "dur": 0.449,
    "args": {
      "Python parent id": 3353, "Python id": 3374, "finished": true, "Ev Idx": 10204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_142", "pid": 537, "tid": 537,
    "ts": 7429564846724.000, "dur": 41.819,
    "args": {
      "Python parent id": 3353, "Python module id": 142, "Python id": 3375, "finished": true, "Ev Idx": 10205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564846724.486, "dur": 41.108,
    "args": {
      "Python parent id": 3375, "Python id": 3376, "finished": true, "Ev Idx": 10206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564846724.678, "dur": 0.241,
    "args": {
      "Python parent id": 3376, "Python id": 3377, "finished": true, "Ev Idx": 10207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564846725.608, "dur": 39.865,
    "args": {
      "Python parent id": 3376, "Python id": 3378, "finished": true, "Ev Idx": 10208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846725.846, "dur": 0.131,
    "args": {
      "Python parent id": 3378, "Python id": 3379, "finished": true, "Ev Idx": 10209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564846726.148, "dur": 0.113,
    "args": {
      "Python parent id": 3378, "Python id": 3380, "finished": true, "Ev Idx": 10210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564846726.378, "dur": 39.003,
    "args": {
      "Python parent id": 3378, "Python id": 3381, "finished": true, "Ev Idx": 10211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564846766.417, "dur": 2.494,
    "args": {
      "Python parent id": 3353, "Python id": 3382, "finished": true, "Ev Idx": 10212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564846769.463, "dur": 2.485,
    "args": {
      "Python parent id": 3353, "Python id": 3383, "finished": true, "Ev Idx": 10213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564846773.191, "dur": 150.451,
    "args": {
      "Python parent id": 3353, "Python id": 3384, "finished": true, "Ev Idx": 10214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564846773.501, "dur": 3.802,
    "args": {
      "Python parent id": 3384, "Python id": 3385, "finished": true, "Ev Idx": 10215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564846777.581, "dur": 1.767,
    "args": {
      "Python parent id": 3384, "Python id": 3386, "finished": true, "Ev Idx": 10216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564846790.080, "dur": 50.808,
    "args": {
      "Python parent id": 3384, "Python id": 3387, "finished": true, "Ev Idx": 10217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564846819.251, "dur": 20.457,
    "args": {
      "Python parent id": 3387, "Python id": 3388, "finished": true, "Ev Idx": 10218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564846872.279, "dur": 29.415,
    "args": {
      "Python parent id": 3384, "Python id": 3389, "finished": true, "Ev Idx": 10219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564846889.968, "dur": 10.943,
    "args": {
      "Python parent id": 3389, "Python id": 3390, "finished": true, "Ev Idx": 10220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564846927.947, "dur": 4.259,
    "args": {
      "Python parent id": 3353, "Python id": 3391, "finished": true, "Ev Idx": 10221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564846928.675, "dur": 0.359,
    "args": {
      "Python parent id": 3391, "Python id": 3392, "finished": true, "Ev Idx": 10222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564846929.422, "dur": 0.088,
    "args": {
      "Python parent id": 3391, "Python id": 3393, "finished": true, "Ev Idx": 10223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564846931.035, "dur": 0.277,
    "args": {
      "Python parent id": 3391, "Python id": 3394, "finished": true, "Ev Idx": 10224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564846931.727, "dur": 0.143,
    "args": {
      "Python parent id": 3391, "Python id": 3395, "finished": true, "Ev Idx": 10225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564846933.689, "dur": 7.159,
    "args": {
      "Python parent id": 3353, "Python id": 3396, "finished": true, "Ev Idx": 10226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564846937.039, "dur": 3.647,
    "args": {
      "Python parent id": 3396, "Python id": 3397, "finished": true, "Ev Idx": 10227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564846937.269, "dur": 1.521,
    "args": {
      "Python parent id": 3397, "Python id": 3398, "finished": true, "Ev Idx": 10228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564846937.793, "dur": 0.844,
    "args": {
      "Python parent id": 3398, "Python id": 3399, "finished": true, "Ev Idx": 10229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564846939.117, "dur": 0.584,
    "args": {
      "Python parent id": 3397, "Python id": 3400, "finished": true, "Ev Idx": 10230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564846940.067, "dur": 0.535,
    "args": {
      "Python parent id": 3397, "Python id": 3401, "finished": true, "Ev Idx": 10231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564846943.592, "dur": 263.282,
    "args": {
      "Python parent id": 3353, "Python id": 3402, "finished": true, "Ev Idx": 10232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564846944.153, "dur": 47.791,
    "args": {
      "Python parent id": 3402, "Python id": 3403, "finished": true, "Ev Idx": 10233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564846963.824, "dur": 5.233,
    "args": {
      "Python parent id": 3403, "Python id": 3404, "finished": true, "Ev Idx": 10234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564846969.893, "dur": 21.906,
    "args": {
      "Python parent id": 3403, "Python id": 3405, "finished": true, "Ev Idx": 10235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564846992.811, "dur": 32.719,
    "args": {
      "Python parent id": 3402, "Python id": 3406, "finished": true, "Ev Idx": 10236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564847008.860, "dur": 2.715,
    "args": {
      "Python parent id": 3406, "Python id": 3407, "finished": true, "Ev Idx": 10237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564847012.083, "dur": 13.357,
    "args": {
      "Python parent id": 3406, "Python id": 3408, "finished": true, "Ev Idx": 10238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564847026.478, "dur": 2.945,
    "args": {
      "Python parent id": 3402, "Python id": 3409, "finished": true, "Ev Idx": 10239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564847029.697, "dur": 49.073,
    "args": {
      "Python parent id": 3402, "Python id": 3410, "finished": true, "Ev Idx": 10240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564847112.378, "dur": 17.704,
    "args": {
      "Python parent id": 3402, "Python id": 3411, "finished": true, "Ev Idx": 10241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564847112.591, "dur": 0.310,
    "args": {
      "Python parent id": 3411, "Python id": 3412, "finished": true, "Ev Idx": 10242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564847113.568, "dur": 16.344,
    "args": {
      "Python parent id": 3411, "Python id": 3413, "finished": true, "Ev Idx": 10243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564847131.034, "dur": 1.907,
    "args": {
      "Python parent id": 3402, "Python id": 3414, "finished": true, "Ev Idx": 10244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564847134.452, "dur": 9.963,
    "args": {
      "Python parent id": 3402, "Python id": 3415, "finished": true, "Ev Idx": 10245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564847134.623, "dur": 0.072,
    "args": {
      "Python parent id": 3415, "Python id": 3416, "finished": true, "Ev Idx": 10246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847140.030, "dur": 0.995,
    "args": {
      "Python parent id": 3415, "Python id": 3417, "finished": true, "Ev Idx": 10247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564847140.496, "dur": 0.449,
    "args": {
      "Python parent id": 3417, "Python id": 3418, "finished": true, "Ev Idx": 10248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564847141.457, "dur": 2.897,
    "args": {
      "Python parent id": 3415, "Python id": 3419, "finished": true, "Ev Idx": 10249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564847145.161, "dur": 41.927,
    "args": {
      "Python parent id": 3402, "Python id": 3420, "finished": true, "Ev Idx": 10250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564847187.753, "dur": 3.462,
    "args": {
      "Python parent id": 3402, "Python id": 3421, "finished": true, "Ev Idx": 10251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564847191.971, "dur": 13.933,
    "args": {
      "Python parent id": 3402, "Python id": 3422, "finished": true, "Ev Idx": 10252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564847209.666, "dur": 3.180,
    "args": {
      "Python parent id": 3353, "Python id": 3423, "finished": true, "Ev Idx": 10253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564847213.588, "dur": 0.239,
    "args": {
      "Python parent id": 3353, "Python id": 3424, "finished": true, "Ev Idx": 10254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847214.895, "dur": 0.840,
    "args": {
      "Python parent id": 3353, "Python id": 3425, "finished": true, "Ev Idx": 10255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_143", "pid": 537, "tid": 537,
    "ts": 7429564847219.932, "dur": 43.727,
    "args": {
      "Python parent id": 3353, "Python module id": 143, "Python id": 3426, "finished": true, "Ev Idx": 10256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847220.929, "dur": 42.398,
    "args": {
      "Python parent id": 3426, "Python id": 3427, "finished": true, "Ev Idx": 10257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847221.379, "dur": 0.833,
    "args": {
      "Python parent id": 3427, "Python id": 3428, "finished": true, "Ev Idx": 10258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564847223.539, "dur": 39.567,
    "args": {
      "Python parent id": 3427, "Python id": 3429, "finished": true, "Ev Idx": 10259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847223.901, "dur": 0.263,
    "args": {
      "Python parent id": 3429, "Python id": 3430, "finished": true, "Ev Idx": 10260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847224.394, "dur": 0.173,
    "args": {
      "Python parent id": 3429, "Python id": 3431, "finished": true, "Ev Idx": 10261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564847224.788, "dur": 38.205,
    "args": {
      "Python parent id": 3429, "Python id": 3432, "finished": true, "Ev Idx": 10262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847278.032, "dur": 1.000,
    "args": {
      "Python parent id": 3337, "Python id": 3433, "finished": true, "Ev Idx": 10263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_41", "pid": 537, "tid": 537,
    "ts": 7429564847281.653, "dur": 139.862,
    "args": {
      "Python parent id": 3337, "Python module id": 41, "Python id": 3434, "finished": true, "Ev Idx": 10264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847282.361, "dur": 138.761,
    "args": {
      "Python parent id": 3434, "Python id": 3435, "finished": true, "Ev Idx": 10265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847282.563, "dur": 0.404,
    "args": {
      "Python parent id": 3435, "Python id": 3436, "finished": true, "Ev Idx": 10266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564847283.799, "dur": 136.303,
    "args": {
      "Python parent id": 3435, "Python id": 3437, "finished": true, "Ev Idx": 10267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564847284.328, "dur": 1.723,
    "args": {
      "Python parent id": 3437, "Python id": 3438, "finished": true, "Ev Idx": 10268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564847286.336, "dur": 13.898,
    "args": {
      "Python parent id": 3437, "Python id": 3439, "finished": true, "Ev Idx": 10269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564847300.586, "dur": 14.819,
    "args": {
      "Python parent id": 3437, "Python id": 3440, "finished": true, "Ev Idx": 10270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564847328.139, "dur": 9.885,
    "args": {
      "Python parent id": 3437, "Python id": 3441, "finished": true, "Ev Idx": 10271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847348.487, "dur": 0.250,
    "args": {
      "Python parent id": 3437, "Python id": 3442, "finished": true, "Ev Idx": 10272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564847349.115, "dur": 1.256,
    "args": {
      "Python parent id": 3437, "Python id": 3443, "finished": true, "Ev Idx": 10273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847422.431, "dur": 0.478,
    "args": {
      "Python parent id": 3337, "Python id": 3444, "finished": true, "Ev Idx": 10274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_20", "pid": 537, "tid": 537,
    "ts": 7429564847424.737, "dur": 236.202,
    "args": {
      "Python parent id": 3337, "Python module id": 20, "Python id": 3445, "finished": true, "Ev Idx": 10275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847425.456, "dur": 235.122,
    "args": {
      "Python parent id": 3445, "Python id": 3446, "finished": true, "Ev Idx": 10276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847425.724, "dur": 0.511,
    "args": {
      "Python parent id": 3446, "Python id": 3447, "finished": true, "Ev Idx": 10277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564847427.409, "dur": 233.045,
    "args": {
      "Python parent id": 3446, "Python id": 3448, "finished": true, "Ev Idx": 10278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847427.704, "dur": 0.356,
    "args": {
      "Python parent id": 3448, "Python id": 3449, "finished": true, "Ev Idx": 10279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847428.463, "dur": 0.305,
    "args": {
      "Python parent id": 3448, "Python id": 3450, "finished": true, "Ev Idx": 10280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847429.046, "dur": 0.203,
    "args": {
      "Python parent id": 3448, "Python id": 3451, "finished": true, "Ev Idx": 10281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_144", "pid": 537, "tid": 537,
    "ts": 7429564847431.716, "dur": 61.641,
    "args": {
      "Python parent id": 3448, "Python module id": 144, "Python id": 3452, "finished": true, "Ev Idx": 10282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847432.208, "dur": 60.871,
    "args": {
      "Python parent id": 3452, "Python id": 3453, "finished": true, "Ev Idx": 10283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847432.346, "dur": 0.187,
    "args": {
      "Python parent id": 3453, "Python id": 3454, "finished": true, "Ev Idx": 10284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564847433.487, "dur": 59.427,
    "args": {
      "Python parent id": 3453, "Python id": 3455, "finished": true, "Ev Idx": 10285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847433.772, "dur": 0.138,
    "args": {
      "Python parent id": 3455, "Python id": 3456, "finished": true, "Ev Idx": 10286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847434.331, "dur": 0.130,
    "args": {
      "Python parent id": 3455, "Python id": 3457, "finished": true, "Ev Idx": 10287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564847434.590, "dur": 58.207,
    "args": {
      "Python parent id": 3455, "Python id": 3458, "finished": true, "Ev Idx": 10288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_20", "pid": 537, "tid": 537,
    "ts": 7429564847495.234, "dur": 17.350,
    "args": {
      "Python parent id": 3448, "Python module id": 20, "Python id": 3459, "finished": true, "Ev Idx": 10289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847495.696, "dur": 16.623,
    "args": {
      "Python parent id": 3459, "Python id": 3460, "finished": true, "Ev Idx": 10290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847495.879, "dur": 0.285,
    "args": {
      "Python parent id": 3460, "Python id": 3461, "finished": true, "Ev Idx": 10291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564847497.391, "dur": 14.789,
    "args": {
      "Python parent id": 3460, "Python id": 3462, "finished": true, "Ev Idx": 10292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564847498.408, "dur": 13.623,
    "args": {
      "Python parent id": 3462, "Python id": 3463, "finished": true, "Ev Idx": 10293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564847498.547, "dur": 0.204,
    "args": {
      "Python parent id": 3463, "Python id": 3464, "finished": true, "Ev Idx": 10294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564847499.709, "dur": 12.192,
    "args": {
      "Python parent id": 3463, "Python id": 3465, "finished": true, "Ev Idx": 10295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847513.356, "dur": 0.428,
    "args": {
      "Python parent id": 3448, "Python id": 3466, "finished": true, "Ev Idx": 10296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_145", "pid": 537, "tid": 537,
    "ts": 7429564847515.589, "dur": 84.890,
    "args": {
      "Python parent id": 3448, "Python module id": 145, "Python id": 3467, "finished": true, "Ev Idx": 10297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847516.018, "dur": 84.239,
    "args": {
      "Python parent id": 3467, "Python id": 3468, "finished": true, "Ev Idx": 10298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847516.206, "dur": 0.232,
    "args": {
      "Python parent id": 3468, "Python id": 3469, "finished": true, "Ev Idx": 10299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564847517.106, "dur": 82.966,
    "args": {
      "Python parent id": 3468, "Python id": 3470, "finished": true, "Ev Idx": 10300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847517.374, "dur": 0.161,
    "args": {
      "Python parent id": 3470, "Python id": 3471, "finished": true, "Ev Idx": 10301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847517.733, "dur": 0.139,
    "args": {
      "Python parent id": 3470, "Python id": 3472, "finished": true, "Ev Idx": 10302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564847517.992, "dur": 81.988,
    "args": {
      "Python parent id": 3470, "Python id": 3473, "finished": true, "Ev Idx": 10303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_146", "pid": 537, "tid": 537,
    "ts": 7429564847613.365, "dur": 46.753,
    "args": {
      "Python parent id": 3448, "Python module id": 146, "Python id": 3474, "finished": true, "Ev Idx": 10304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847614.064, "dur": 45.816,
    "args": {
      "Python parent id": 3474, "Python id": 3475, "finished": true, "Ev Idx": 10305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847614.279, "dur": 0.385,
    "args": {
      "Python parent id": 3475, "Python id": 3476, "finished": true, "Ev Idx": 10306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564847615.509, "dur": 44.180,
    "args": {
      "Python parent id": 3475, "Python id": 3477, "finished": true, "Ev Idx": 10307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847615.808, "dur": 0.212,
    "args": {
      "Python parent id": 3477, "Python id": 3478, "finished": true, "Ev Idx": 10308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847616.230, "dur": 0.125,
    "args": {
      "Python parent id": 3477, "Python id": 3479, "finished": true, "Ev Idx": 10309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564847616.473, "dur": 43.162,
    "args": {
      "Python parent id": 3477, "Python id": 3480, "finished": true, "Ev Idx": 10310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_21", "pid": 537, "tid": 537,
    "ts": 7429564847680.422, "dur": 1205.511,
    "args": {
      "Python parent id": 64, "Python module id": 21, "Python id": 3481, "finished": true, "Ev Idx": 10311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847681.950, "dur": 1203.091,
    "args": {
      "Python parent id": 3481, "Python id": 3482, "finished": true, "Ev Idx": 10312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847682.165, "dur": 0.356,
    "args": {
      "Python parent id": 3482, "Python id": 3483, "finished": true, "Ev Idx": 10313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564847684.194, "dur": 1199.839,
    "args": {
      "Python parent id": 3482, "Python id": 3484, "finished": true, "Ev Idx": 10314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847684.707, "dur": 0.554,
    "args": {
      "Python parent id": 3484, "Python id": 3485, "finished": true, "Ev Idx": 10315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_42", "pid": 537, "tid": 537,
    "ts": 7429564847687.369, "dur": 78.772,
    "args": {
      "Python parent id": 3484, "Python module id": 42, "Python id": 3486, "finished": true, "Ev Idx": 10316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847687.888, "dur": 77.907,
    "args": {
      "Python parent id": 3486, "Python id": 3487, "finished": true, "Ev Idx": 10317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847688.061, "dur": 0.162,
    "args": {
      "Python parent id": 3487, "Python id": 3488, "finished": true, "Ev Idx": 10318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564847689.111, "dur": 75.955,
    "args": {
      "Python parent id": 3487, "Python id": 3489, "finished": true, "Ev Idx": 10319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564847689.681, "dur": 1.909,
    "args": {
      "Python parent id": 3489, "Python id": 3490, "finished": true, "Ev Idx": 10320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564847691.875, "dur": 11.908,
    "args": {
      "Python parent id": 3489, "Python id": 3491, "finished": true, "Ev Idx": 10321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564847704.089, "dur": 15.487,
    "args": {
      "Python parent id": 3489, "Python id": 3492, "finished": true, "Ev Idx": 10322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564847731.094, "dur": 11.289,
    "args": {
      "Python parent id": 3489, "Python id": 3493, "finished": true, "Ev Idx": 10323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847753.344, "dur": 0.216,
    "args": {
      "Python parent id": 3489, "Python id": 3494, "finished": true, "Ev Idx": 10324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564847753.907, "dur": 1.091,
    "args": {
      "Python parent id": 3489, "Python id": 3495, "finished": true, "Ev Idx": 10325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847766.718, "dur": 0.379,
    "args": {
      "Python parent id": 3484, "Python id": 3496, "finished": true, "Ev Idx": 10326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_21", "pid": 537, "tid": 537,
    "ts": 7429564847769.616, "dur": 715.025,
    "args": {
      "Python parent id": 3484, "Python module id": 21, "Python id": 3497, "finished": true, "Ev Idx": 10327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847770.976, "dur": 712.990,
    "args": {
      "Python parent id": 3497, "Python id": 3498, "finished": true, "Ev Idx": 10328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847771.214, "dur": 0.244,
    "args": {
      "Python parent id": 3498, "Python id": 3499, "finished": true, "Ev Idx": 10329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564847773.007, "dur": 710.286,
    "args": {
      "Python parent id": 3498, "Python id": 3500, "finished": true, "Ev Idx": 10330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847776.221, "dur": 0.460,
    "args": {
      "Python parent id": 3500, "Python id": 3501, "finished": true, "Ev Idx": 10331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_147", "pid": 537, "tid": 537,
    "ts": 7429564847778.421, "dur": 81.420,
    "args": {
      "Python parent id": 3500, "Python module id": 147, "Python id": 3502, "finished": true, "Ev Idx": 10332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847779.078, "dur": 80.410,
    "args": {
      "Python parent id": 3502, "Python id": 3503, "finished": true, "Ev Idx": 10333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847779.228, "dur": 0.185,
    "args": {
      "Python parent id": 3503, "Python id": 3504, "finished": true, "Ev Idx": 10334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564847780.367, "dur": 78.910,
    "args": {
      "Python parent id": 3503, "Python id": 3505, "finished": true, "Ev Idx": 10335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847780.672, "dur": 0.151,
    "args": {
      "Python parent id": 3505, "Python id": 3506, "finished": true, "Ev Idx": 10336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847781.016, "dur": 0.146,
    "args": {
      "Python parent id": 3505, "Python id": 3507, "finished": true, "Ev Idx": 10337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564847781.306, "dur": 77.861,
    "args": {
      "Python parent id": 3505, "Python id": 3508, "finished": true, "Ev Idx": 10338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564847860.465, "dur": 3.620,
    "args": {
      "Python parent id": 3500, "Python id": 3509, "finished": true, "Ev Idx": 10339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564847864.642, "dur": 3.828,
    "args": {
      "Python parent id": 3500, "Python id": 3510, "finished": true, "Ev Idx": 10340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847869.361, "dur": 0.462,
    "args": {
      "Python parent id": 3500, "Python id": 3511, "finished": true, "Ev Idx": 10341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_148", "pid": 537, "tid": 537,
    "ts": 7429564847871.962, "dur": 45.753,
    "args": {
      "Python parent id": 3500, "Python module id": 148, "Python id": 3512, "finished": true, "Ev Idx": 10342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847872.518, "dur": 44.972,
    "args": {
      "Python parent id": 3512, "Python id": 3513, "finished": true, "Ev Idx": 10343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847872.806, "dur": 0.312,
    "args": {
      "Python parent id": 3513, "Python id": 3514, "finished": true, "Ev Idx": 10344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564847873.899, "dur": 43.487,
    "args": {
      "Python parent id": 3513, "Python id": 3515, "finished": true, "Ev Idx": 10345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847874.164, "dur": 0.132,
    "args": {
      "Python parent id": 3515, "Python id": 3516, "finished": true, "Ev Idx": 10346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847874.505, "dur": 0.136,
    "args": {
      "Python parent id": 3515, "Python id": 3517, "finished": true, "Ev Idx": 10347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564847874.783, "dur": 42.523,
    "args": {
      "Python parent id": 3515, "Python id": 3518, "finished": true, "Ev Idx": 10348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564847918.241, "dur": 2.408,
    "args": {
      "Python parent id": 3500, "Python id": 3519, "finished": true, "Ev Idx": 10349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564847921.096, "dur": 2.684,
    "args": {
      "Python parent id": 3500, "Python id": 3520, "finished": true, "Ev Idx": 10350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847924.494, "dur": 0.463,
    "args": {
      "Python parent id": 3500, "Python id": 3521, "finished": true, "Ev Idx": 10351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_149", "pid": 537, "tid": 537,
    "ts": 7429564847927.469, "dur": 43.876,
    "args": {
      "Python parent id": 3500, "Python module id": 149, "Python id": 3522, "finished": true, "Ev Idx": 10352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564847928.089, "dur": 43.042,
    "args": {
      "Python parent id": 3522, "Python id": 3523, "finished": true, "Ev Idx": 10353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564847928.290, "dur": 0.258,
    "args": {
      "Python parent id": 3523, "Python id": 3524, "finished": true, "Ev Idx": 10354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564847929.227, "dur": 41.799,
    "args": {
      "Python parent id": 3523, "Python id": 3525, "finished": true, "Ev Idx": 10355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847929.397, "dur": 0.135,
    "args": {
      "Python parent id": 3525, "Python id": 3526, "finished": true, "Ev Idx": 10356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564847929.710, "dur": 0.115,
    "args": {
      "Python parent id": 3525, "Python id": 3527, "finished": true, "Ev Idx": 10357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564847929.931, "dur": 41.020,
    "args": {
      "Python parent id": 3525, "Python id": 3528, "finished": true, "Ev Idx": 10358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564847971.888, "dur": 2.487,
    "args": {
      "Python parent id": 3500, "Python id": 3529, "finished": true, "Ev Idx": 10359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564847974.955, "dur": 2.626,
    "args": {
      "Python parent id": 3500, "Python id": 3530, "finished": true, "Ev Idx": 10360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564847978.710, "dur": 149.218,
    "args": {
      "Python parent id": 3500, "Python id": 3531, "finished": true, "Ev Idx": 10361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564847978.986, "dur": 3.595,
    "args": {
      "Python parent id": 3531, "Python id": 3532, "finished": true, "Ev Idx": 10362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564847982.842, "dur": 1.882,
    "args": {
      "Python parent id": 3531, "Python id": 3533, "finished": true, "Ev Idx": 10363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564847997.716, "dur": 49.010,
    "args": {
      "Python parent id": 3531, "Python id": 3534, "finished": true, "Ev Idx": 10364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564848025.363, "dur": 20.184,
    "args": {
      "Python parent id": 3534, "Python id": 3535, "finished": true, "Ev Idx": 10365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564848078.618, "dur": 30.893,
    "args": {
      "Python parent id": 3531, "Python id": 3536, "finished": true, "Ev Idx": 10366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564848096.097, "dur": 12.824,
    "args": {
      "Python parent id": 3536, "Python id": 3537, "finished": true, "Ev Idx": 10367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564848132.218, "dur": 3.940,
    "args": {
      "Python parent id": 3500, "Python id": 3538, "finished": true, "Ev Idx": 10368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564848132.891, "dur": 0.367,
    "args": {
      "Python parent id": 3538, "Python id": 3539, "finished": true, "Ev Idx": 10369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564848133.633, "dur": 0.085,
    "args": {
      "Python parent id": 3538, "Python id": 3540, "finished": true, "Ev Idx": 10370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564848135.105, "dur": 0.251,
    "args": {
      "Python parent id": 3538, "Python id": 3541, "finished": true, "Ev Idx": 10371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564848135.819, "dur": 0.111,
    "args": {
      "Python parent id": 3538, "Python id": 3542, "finished": true, "Ev Idx": 10372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564848137.596, "dur": 6.588,
    "args": {
      "Python parent id": 3500, "Python id": 3543, "finished": true, "Ev Idx": 10373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564848140.657, "dur": 3.376,
    "args": {
      "Python parent id": 3543, "Python id": 3544, "finished": true, "Ev Idx": 10374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564848140.878, "dur": 1.579,
    "args": {
      "Python parent id": 3544, "Python id": 3545, "finished": true, "Ev Idx": 10375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564848141.404, "dur": 0.854,
    "args": {
      "Python parent id": 3545, "Python id": 3546, "finished": true, "Ev Idx": 10376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564848142.771, "dur": 0.488,
    "args": {
      "Python parent id": 3544, "Python id": 3547, "finished": true, "Ev Idx": 10377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564848143.544, "dur": 0.416,
    "args": {
      "Python parent id": 3544, "Python id": 3548, "finished": true, "Ev Idx": 10378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564848146.760, "dur": 282.203,
    "args": {
      "Python parent id": 3500, "Python id": 3549, "finished": true, "Ev Idx": 10379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564848147.300, "dur": 45.901,
    "args": {
      "Python parent id": 3549, "Python id": 3550, "finished": true, "Ev Idx": 10380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564848165.503, "dur": 5.229,
    "args": {
      "Python parent id": 3550, "Python id": 3551, "finished": true, "Ev Idx": 10381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564848171.585, "dur": 21.485,
    "args": {
      "Python parent id": 3550, "Python id": 3552, "finished": true, "Ev Idx": 10382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564848194.009, "dur": 63.332,
    "args": {
      "Python parent id": 3549, "Python id": 3553, "finished": true, "Ev Idx": 10383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564848207.844, "dur": 3.544,
    "args": {
      "Python parent id": 3553, "Python id": 3554, "finished": true, "Ev Idx": 10384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564848211.942, "dur": 45.339,
    "args": {
      "Python parent id": 3553, "Python id": 3555, "finished": true, "Ev Idx": 10385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564848258.287, "dur": 2.681,
    "args": {
      "Python parent id": 3549, "Python id": 3556, "finished": true, "Ev Idx": 10386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564848261.282, "dur": 46.760,
    "args": {
      "Python parent id": 3549, "Python id": 3557, "finished": true, "Ev Idx": 10387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564848339.418, "dur": 16.424,
    "args": {
      "Python parent id": 3549, "Python id": 3558, "finished": true, "Ev Idx": 10388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564848339.588, "dur": 0.247,
    "args": {
      "Python parent id": 3558, "Python id": 3559, "finished": true, "Ev Idx": 10389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564848340.508, "dur": 15.168,
    "args": {
      "Python parent id": 3558, "Python id": 3560, "finished": true, "Ev Idx": 10390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564848356.798, "dur": 1.887,
    "args": {
      "Python parent id": 3549, "Python id": 3561, "finished": true, "Ev Idx": 10391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564848360.312, "dur": 9.569,
    "args": {
      "Python parent id": 3549, "Python id": 3562, "finished": true, "Ev Idx": 10392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564848360.423, "dur": 0.118,
    "args": {
      "Python parent id": 3562, "Python id": 3563, "finished": true, "Ev Idx": 10393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848365.545, "dur": 1.039,
    "args": {
      "Python parent id": 3562, "Python id": 3564, "finished": true, "Ev Idx": 10394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564848366.059, "dur": 0.447,
    "args": {
      "Python parent id": 3564, "Python id": 3565, "finished": true, "Ev Idx": 10395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564848366.994, "dur": 2.821,
    "args": {
      "Python parent id": 3562, "Python id": 3566, "finished": true, "Ev Idx": 10396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564848370.661, "dur": 38.446,
    "args": {
      "Python parent id": 3549, "Python id": 3567, "finished": true, "Ev Idx": 10397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564848409.768, "dur": 3.029,
    "args": {
      "Python parent id": 3549, "Python id": 3568, "finished": true, "Ev Idx": 10398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564848413.535, "dur": 14.357,
    "args": {
      "Python parent id": 3549, "Python id": 3569, "finished": true, "Ev Idx": 10399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564848431.711, "dur": 3.080,
    "args": {
      "Python parent id": 3500, "Python id": 3570, "finished": true, "Ev Idx": 10400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564848435.498, "dur": 0.272,
    "args": {
      "Python parent id": 3500, "Python id": 3571, "finished": true, "Ev Idx": 10401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848436.750, "dur": 0.741,
    "args": {
      "Python parent id": 3500, "Python id": 3572, "finished": true, "Ev Idx": 10402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_150", "pid": 537, "tid": 537,
    "ts": 7429564848441.243, "dur": 41.568,
    "args": {
      "Python parent id": 3500, "Python module id": 150, "Python id": 3573, "finished": true, "Ev Idx": 10403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848442.046, "dur": 40.490,
    "args": {
      "Python parent id": 3573, "Python id": 3574, "finished": true, "Ev Idx": 10404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848442.523, "dur": 0.933,
    "args": {
      "Python parent id": 3574, "Python id": 3575, "finished": true, "Ev Idx": 10405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564848444.722, "dur": 37.553,
    "args": {
      "Python parent id": 3574, "Python id": 3576, "finished": true, "Ev Idx": 10406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848445.141, "dur": 0.211,
    "args": {
      "Python parent id": 3576, "Python id": 3577, "finished": true, "Ev Idx": 10407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848445.648, "dur": 0.161,
    "args": {
      "Python parent id": 3576, "Python id": 3578, "finished": true, "Ev Idx": 10408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564848445.939, "dur": 36.261,
    "args": {
      "Python parent id": 3576, "Python id": 3579, "finished": true, "Ev Idx": 10409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848496.904, "dur": 0.606,
    "args": {
      "Python parent id": 3484, "Python id": 3580, "finished": true, "Ev Idx": 10410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_43", "pid": 537, "tid": 537,
    "ts": 7429564848500.008, "dur": 82.171,
    "args": {
      "Python parent id": 3484, "Python module id": 43, "Python id": 3581, "finished": true, "Ev Idx": 10411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848500.580, "dur": 81.258,
    "args": {
      "Python parent id": 3581, "Python id": 3582, "finished": true, "Ev Idx": 10412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848500.812, "dur": 0.385,
    "args": {
      "Python parent id": 3582, "Python id": 3583, "finished": true, "Ev Idx": 10413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564848502.729, "dur": 78.223,
    "args": {
      "Python parent id": 3582, "Python id": 3584, "finished": true, "Ev Idx": 10414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564848503.239, "dur": 1.478,
    "args": {
      "Python parent id": 3584, "Python id": 3585, "finished": true, "Ev Idx": 10415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564848504.996, "dur": 12.456,
    "args": {
      "Python parent id": 3584, "Python id": 3586, "finished": true, "Ev Idx": 10416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564848517.849, "dur": 16.180,
    "args": {
      "Python parent id": 3584, "Python id": 3587, "finished": true, "Ev Idx": 10417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564848547.773, "dur": 11.866,
    "args": {
      "Python parent id": 3584, "Python id": 3588, "finished": true, "Ev Idx": 10418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848569.332, "dur": 0.211,
    "args": {
      "Python parent id": 3584, "Python id": 3589, "finished": true, "Ev Idx": 10419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564848569.935, "dur": 1.180,
    "args": {
      "Python parent id": 3584, "Python id": 3590, "finished": true, "Ev Idx": 10420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848582.852, "dur": 0.409,
    "args": {
      "Python parent id": 3484, "Python id": 3591, "finished": true, "Ev Idx": 10421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_21", "pid": 537, "tid": 537,
    "ts": 7429564848585.247, "dur": 286.534,
    "args": {
      "Python parent id": 3484, "Python module id": 21, "Python id": 3592, "finished": true, "Ev Idx": 10422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848585.809, "dur": 285.576,
    "args": {
      "Python parent id": 3592, "Python id": 3593, "finished": true, "Ev Idx": 10423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848586.005, "dur": 0.282,
    "args": {
      "Python parent id": 3593, "Python id": 3594, "finished": true, "Ev Idx": 10424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564848587.217, "dur": 284.049,
    "args": {
      "Python parent id": 3593, "Python id": 3595, "finished": true, "Ev Idx": 10425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848587.554, "dur": 1.774,
    "args": {
      "Python parent id": 3595, "Python id": 3596, "finished": true, "Ev Idx": 10426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848589.664, "dur": 0.316,
    "args": {
      "Python parent id": 3595, "Python id": 3597, "finished": true, "Ev Idx": 10427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848590.251, "dur": 0.205,
    "args": {
      "Python parent id": 3595, "Python id": 3598, "finished": true, "Ev Idx": 10428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_151", "pid": 537, "tid": 537,
    "ts": 7429564848592.310, "dur": 112.943,
    "args": {
      "Python parent id": 3595, "Python module id": 151, "Python id": 3599, "finished": true, "Ev Idx": 10429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848592.981, "dur": 111.985,
    "args": {
      "Python parent id": 3599, "Python id": 3600, "finished": true, "Ev Idx": 10430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848593.130, "dur": 0.151,
    "args": {
      "Python parent id": 3600, "Python id": 3601, "finished": true, "Ev Idx": 10431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564848593.948, "dur": 110.846,
    "args": {
      "Python parent id": 3600, "Python id": 3602, "finished": true, "Ev Idx": 10432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848594.298, "dur": 0.148,
    "args": {
      "Python parent id": 3602, "Python id": 3603, "finished": true, "Ev Idx": 10433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848594.666, "dur": 0.115,
    "args": {
      "Python parent id": 3602, "Python id": 3604, "finished": true, "Ev Idx": 10434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564848594.890, "dur": 109.816,
    "args": {
      "Python parent id": 3602, "Python id": 3605, "finished": true, "Ev Idx": 10435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_21", "pid": 537, "tid": 537,
    "ts": 7429564848707.602, "dur": 17.297,
    "args": {
      "Python parent id": 3595, "Python module id": 21, "Python id": 3606, "finished": true, "Ev Idx": 10436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848708.166, "dur": 16.486,
    "args": {
      "Python parent id": 3606, "Python id": 3607, "finished": true, "Ev Idx": 10437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848708.417, "dur": 0.499,
    "args": {
      "Python parent id": 3607, "Python id": 3608, "finished": true, "Ev Idx": 10438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564848709.943, "dur": 14.582,
    "args": {
      "Python parent id": 3607, "Python id": 3609, "finished": true, "Ev Idx": 10439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564848710.903, "dur": 13.470,
    "args": {
      "Python parent id": 3609, "Python id": 3610, "finished": true, "Ev Idx": 10440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564848711.056, "dur": 0.175,
    "args": {
      "Python parent id": 3610, "Python id": 3611, "finished": true, "Ev Idx": 10441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564848712.183, "dur": 12.094,
    "args": {
      "Python parent id": 3610, "Python id": 3612, "finished": true, "Ev Idx": 10442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848725.723, "dur": 0.402,
    "args": {
      "Python parent id": 3595, "Python id": 3613, "finished": true, "Ev Idx": 10443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_152", "pid": 537, "tid": 537,
    "ts": 7429564848728.195, "dur": 82.565,
    "args": {
      "Python parent id": 3595, "Python module id": 152, "Python id": 3614, "finished": true, "Ev Idx": 10444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848728.658, "dur": 81.808,
    "args": {
      "Python parent id": 3614, "Python id": 3615, "finished": true, "Ev Idx": 10445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848728.816, "dur": 0.190,
    "args": {
      "Python parent id": 3615, "Python id": 3616, "finished": true, "Ev Idx": 10446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564848729.844, "dur": 80.401,
    "args": {
      "Python parent id": 3615, "Python id": 3617, "finished": true, "Ev Idx": 10447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848730.131, "dur": 0.179,
    "args": {
      "Python parent id": 3617, "Python id": 3618, "finished": true, "Ev Idx": 10448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848730.498, "dur": 0.128,
    "args": {
      "Python parent id": 3617, "Python id": 3619, "finished": true, "Ev Idx": 10449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564848730.733, "dur": 79.424,
    "args": {
      "Python parent id": 3617, "Python id": 3620, "finished": true, "Ev Idx": 10450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_153", "pid": 537, "tid": 537,
    "ts": 7429564848824.512, "dur": 46.466,
    "args": {
      "Python parent id": 3595, "Python module id": 153, "Python id": 3621, "finished": true, "Ev Idx": 10451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848825.190, "dur": 45.590,
    "args": {
      "Python parent id": 3621, "Python id": 3622, "finished": true, "Ev Idx": 10452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848825.558, "dur": 0.493,
    "args": {
      "Python parent id": 3622, "Python id": 3623, "finished": true, "Ev Idx": 10453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564848826.964, "dur": 43.724,
    "args": {
      "Python parent id": 3622, "Python id": 3624, "finished": true, "Ev Idx": 10454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848827.311, "dur": 0.272,
    "args": {
      "Python parent id": 3624, "Python id": 3625, "finished": true, "Ev Idx": 10455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848827.806, "dur": 0.140,
    "args": {
      "Python parent id": 3624, "Python id": 3626, "finished": true, "Ev Idx": 10456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564848828.091, "dur": 42.549,
    "args": {
      "Python parent id": 3624, "Python id": 3627, "finished": true, "Ev Idx": 10457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_22", "pid": 537, "tid": 537,
    "ts": 7429564848892.715, "dur": 1226.832,
    "args": {
      "Python parent id": 64, "Python module id": 22, "Python id": 3628, "finished": true, "Ev Idx": 10458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848894.135, "dur": 1224.607,
    "args": {
      "Python parent id": 3628, "Python id": 3629, "finished": true, "Ev Idx": 10459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848894.372, "dur": 0.337,
    "args": {
      "Python parent id": 3629, "Python id": 3630, "finished": true, "Ev Idx": 10460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564848896.506, "dur": 1221.277,
    "args": {
      "Python parent id": 3629, "Python id": 3631, "finished": true, "Ev Idx": 10461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848897.021, "dur": 0.578,
    "args": {
      "Python parent id": 3631, "Python id": 3632, "finished": true, "Ev Idx": 10462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_44", "pid": 537, "tid": 537,
    "ts": 7429564848899.600, "dur": 79.931,
    "args": {
      "Python parent id": 3631, "Python module id": 44, "Python id": 3633, "finished": true, "Ev Idx": 10463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848900.178, "dur": 79.001,
    "args": {
      "Python parent id": 3633, "Python id": 3634, "finished": true, "Ev Idx": 10464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848900.326, "dur": 0.175,
    "args": {
      "Python parent id": 3634, "Python id": 3635, "finished": true, "Ev Idx": 10465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564848901.341, "dur": 77.095,
    "args": {
      "Python parent id": 3634, "Python id": 3636, "finished": true, "Ev Idx": 10466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564848901.920, "dur": 1.903,
    "args": {
      "Python parent id": 3636, "Python id": 3637, "finished": true, "Ev Idx": 10467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564848904.146, "dur": 13.405,
    "args": {
      "Python parent id": 3636, "Python id": 3638, "finished": true, "Ev Idx": 10468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564848917.935, "dur": 14.970,
    "args": {
      "Python parent id": 3636, "Python id": 3639, "finished": true, "Ev Idx": 10469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564848944.358, "dur": 10.445,
    "args": {
      "Python parent id": 3636, "Python id": 3640, "finished": true, "Ev Idx": 10470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848965.139, "dur": 0.193,
    "args": {
      "Python parent id": 3636, "Python id": 3641, "finished": true, "Ev Idx": 10471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564848965.833, "dur": 1.342,
    "args": {
      "Python parent id": 3636, "Python id": 3642, "finished": true, "Ev Idx": 10472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848980.239, "dur": 0.486,
    "args": {
      "Python parent id": 3631, "Python id": 3643, "finished": true, "Ev Idx": 10473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_22", "pid": 537, "tid": 537,
    "ts": 7429564848983.529, "dur": 719.061,
    "args": {
      "Python parent id": 3631, "Python module id": 22, "Python id": 3644, "finished": true, "Ev Idx": 10474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848985.030, "dur": 716.753,
    "args": {
      "Python parent id": 3644, "Python id": 3645, "finished": true, "Ev Idx": 10475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848985.315, "dur": 0.325,
    "args": {
      "Python parent id": 3645, "Python id": 3646, "finished": true, "Ev Idx": 10476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564848987.243, "dur": 713.846,
    "args": {
      "Python parent id": 3645, "Python id": 3647, "finished": true, "Ev Idx": 10477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848990.618, "dur": 0.389,
    "args": {
      "Python parent id": 3647, "Python id": 3648, "finished": true, "Ev Idx": 10478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_154", "pid": 537, "tid": 537,
    "ts": 7429564848993.028, "dur": 44.602,
    "args": {
      "Python parent id": 3647, "Python module id": 154, "Python id": 3649, "finished": true, "Ev Idx": 10479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564848993.585, "dur": 43.789,
    "args": {
      "Python parent id": 3649, "Python id": 3650, "finished": true, "Ev Idx": 10480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564848993.738, "dur": 0.172,
    "args": {
      "Python parent id": 3650, "Python id": 3651, "finished": true, "Ev Idx": 10481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564848994.564, "dur": 42.698,
    "args": {
      "Python parent id": 3650, "Python id": 3652, "finished": true, "Ev Idx": 10482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848994.953, "dur": 0.142,
    "args": {
      "Python parent id": 3652, "Python id": 3653, "finished": true, "Ev Idx": 10483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564848995.302, "dur": 0.141,
    "args": {
      "Python parent id": 3652, "Python id": 3654, "finished": true, "Ev Idx": 10484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564848995.611, "dur": 41.600,
    "args": {
      "Python parent id": 3652, "Python id": 3655, "finished": true, "Ev Idx": 10485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564849038.344, "dur": 3.396,
    "args": {
      "Python parent id": 3647, "Python id": 3656, "finished": true, "Ev Idx": 10486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564849042.284, "dur": 4.350,
    "args": {
      "Python parent id": 3647, "Python id": 3657, "finished": true, "Ev Idx": 10487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849047.433, "dur": 0.351,
    "args": {
      "Python parent id": 3647, "Python id": 3658, "finished": true, "Ev Idx": 10488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_155", "pid": 537, "tid": 537,
    "ts": 7429564849049.854, "dur": 48.010,
    "args": {
      "Python parent id": 3647, "Python module id": 155, "Python id": 3659, "finished": true, "Ev Idx": 10489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564849050.592, "dur": 47.053,
    "args": {
      "Python parent id": 3659, "Python id": 3660, "finished": true, "Ev Idx": 10490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564849050.822, "dur": 0.307,
    "args": {
      "Python parent id": 3660, "Python id": 3661, "finished": true, "Ev Idx": 10491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564849051.881, "dur": 45.661,
    "args": {
      "Python parent id": 3660, "Python id": 3662, "finished": true, "Ev Idx": 10492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849052.117, "dur": 0.125,
    "args": {
      "Python parent id": 3662, "Python id": 3663, "finished": true, "Ev Idx": 10493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849052.445, "dur": 0.129,
    "args": {
      "Python parent id": 3662, "Python id": 3664, "finished": true, "Ev Idx": 10494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564849052.693, "dur": 44.772,
    "args": {
      "Python parent id": 3662, "Python id": 3665, "finished": true, "Ev Idx": 10495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564849098.326, "dur": 2.535,
    "args": {
      "Python parent id": 3647, "Python id": 3666, "finished": true, "Ev Idx": 10496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564849101.353, "dur": 3.067,
    "args": {
      "Python parent id": 3647, "Python id": 3667, "finished": true, "Ev Idx": 10497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849105.154, "dur": 0.463,
    "args": {
      "Python parent id": 3647, "Python id": 3668, "finished": true, "Ev Idx": 10498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_156", "pid": 537, "tid": 537,
    "ts": 7429564849107.699, "dur": 41.744,
    "args": {
      "Python parent id": 3647, "Python module id": 156, "Python id": 3669, "finished": true, "Ev Idx": 10499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564849108.159, "dur": 41.104,
    "args": {
      "Python parent id": 3669, "Python id": 3670, "finished": true, "Ev Idx": 10500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564849108.319, "dur": 0.245,
    "args": {
      "Python parent id": 3670, "Python id": 3671, "finished": true, "Ev Idx": 10501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564849109.204, "dur": 39.941,
    "args": {
      "Python parent id": 3670, "Python id": 3672, "finished": true, "Ev Idx": 10502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849109.372, "dur": 0.147,
    "args": {
      "Python parent id": 3672, "Python id": 3673, "finished": true, "Ev Idx": 10503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849109.685, "dur": 0.119,
    "args": {
      "Python parent id": 3672, "Python id": 3674, "finished": true, "Ev Idx": 10504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564849109.921, "dur": 39.145,
    "args": {
      "Python parent id": 3672, "Python id": 3675, "finished": true, "Ev Idx": 10505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564849149.965, "dur": 2.474,
    "args": {
      "Python parent id": 3647, "Python id": 3676, "finished": true, "Ev Idx": 10506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564849152.943, "dur": 2.569,
    "args": {
      "Python parent id": 3647, "Python id": 3677, "finished": true, "Ev Idx": 10507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564849156.765, "dur": 200.279,
    "args": {
      "Python parent id": 3647, "Python id": 3678, "finished": true, "Ev Idx": 10508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564849157.007, "dur": 3.854,
    "args": {
      "Python parent id": 3678, "Python id": 3679, "finished": true, "Ev Idx": 10509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564849161.159, "dur": 1.908,
    "args": {
      "Python parent id": 3678, "Python id": 3680, "finished": true, "Ev Idx": 10510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564849175.104, "dur": 53.894,
    "args": {
      "Python parent id": 3678, "Python id": 3681, "finished": true, "Ev Idx": 10511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564849205.413, "dur": 22.148,
    "args": {
      "Python parent id": 3681, "Python id": 3682, "finished": true, "Ev Idx": 10512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564849305.114, "dur": 33.685,
    "args": {
      "Python parent id": 3678, "Python id": 3683, "finished": true, "Ev Idx": 10513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564849322.877, "dur": 15.048,
    "args": {
      "Python parent id": 3683, "Python id": 3684, "finished": true, "Ev Idx": 10514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564849361.722, "dur": 4.138,
    "args": {
      "Python parent id": 3647, "Python id": 3685, "finished": true, "Ev Idx": 10515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564849362.537, "dur": 0.396,
    "args": {
      "Python parent id": 3685, "Python id": 3686, "finished": true, "Ev Idx": 10516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564849363.279, "dur": 0.074,
    "args": {
      "Python parent id": 3685, "Python id": 3687, "finished": true, "Ev Idx": 10517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564849364.775, "dur": 0.278,
    "args": {
      "Python parent id": 3685, "Python id": 3688, "finished": true, "Ev Idx": 10518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564849365.455, "dur": 0.142,
    "args": {
      "Python parent id": 3685, "Python id": 3689, "finished": true, "Ev Idx": 10519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564849367.459, "dur": 6.983,
    "args": {
      "Python parent id": 3647, "Python id": 3690, "finished": true, "Ev Idx": 10520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564849370.741, "dur": 3.539,
    "args": {
      "Python parent id": 3690, "Python id": 3691, "finished": true, "Ev Idx": 10521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564849370.972, "dur": 1.594,
    "args": {
      "Python parent id": 3691, "Python id": 3692, "finished": true, "Ev Idx": 10522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564849371.455, "dur": 0.923,
    "args": {
      "Python parent id": 3692, "Python id": 3693, "finished": true, "Ev Idx": 10523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564849372.850, "dur": 0.558,
    "args": {
      "Python parent id": 3691, "Python id": 3694, "finished": true, "Ev Idx": 10524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564849373.775, "dur": 0.413,
    "args": {
      "Python parent id": 3691, "Python id": 3695, "finished": true, "Ev Idx": 10525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564849377.441, "dur": 265.192,
    "args": {
      "Python parent id": 3647, "Python id": 3696, "finished": true, "Ev Idx": 10526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564849377.993, "dur": 49.275,
    "args": {
      "Python parent id": 3696, "Python id": 3697, "finished": true, "Ev Idx": 10527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564849397.848, "dur": 6.404,
    "args": {
      "Python parent id": 3697, "Python id": 3698, "finished": true, "Ev Idx": 10528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564849405.108, "dur": 22.006,
    "args": {
      "Python parent id": 3697, "Python id": 3699, "finished": true, "Ev Idx": 10529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564849428.215, "dur": 32.804,
    "args": {
      "Python parent id": 3696, "Python id": 3700, "finished": true, "Ev Idx": 10530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564849443.916, "dur": 3.304,
    "args": {
      "Python parent id": 3700, "Python id": 3701, "finished": true, "Ev Idx": 10531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564849447.660, "dur": 13.267,
    "args": {
      "Python parent id": 3700, "Python id": 3702, "finished": true, "Ev Idx": 10532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564849462.074, "dur": 3.958,
    "args": {
      "Python parent id": 3696, "Python id": 3703, "finished": true, "Ev Idx": 10533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564849466.391, "dur": 48.715,
    "args": {
      "Python parent id": 3696, "Python id": 3704, "finished": true, "Ev Idx": 10534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564849548.101, "dur": 16.644,
    "args": {
      "Python parent id": 3696, "Python id": 3705, "finished": true, "Ev Idx": 10535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564849548.316, "dur": 0.360,
    "args": {
      "Python parent id": 3705, "Python id": 3706, "finished": true, "Ev Idx": 10536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564849549.283, "dur": 15.300,
    "args": {
      "Python parent id": 3705, "Python id": 3707, "finished": true, "Ev Idx": 10537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564849565.721, "dur": 1.893,
    "args": {
      "Python parent id": 3696, "Python id": 3708, "finished": true, "Ev Idx": 10538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564849569.343, "dur": 10.116,
    "args": {
      "Python parent id": 3696, "Python id": 3709, "finished": true, "Ev Idx": 10539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564849569.593, "dur": 0.134,
    "args": {
      "Python parent id": 3709, "Python id": 3710, "finished": true, "Ev Idx": 10540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849575.191, "dur": 1.039,
    "args": {
      "Python parent id": 3709, "Python id": 3711, "finished": true, "Ev Idx": 10541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564849575.692, "dur": 0.446,
    "args": {
      "Python parent id": 3711, "Python id": 3712, "finished": true, "Ev Idx": 10542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564849576.640, "dur": 2.724,
    "args": {
      "Python parent id": 3709, "Python id": 3713, "finished": true, "Ev Idx": 10543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564849580.211, "dur": 42.191,
    "args": {
      "Python parent id": 3696, "Python id": 3714, "finished": true, "Ev Idx": 10544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564849623.032, "dur": 3.330,
    "args": {
      "Python parent id": 3696, "Python id": 3715, "finished": true, "Ev Idx": 10545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564849627.014, "dur": 14.614,
    "args": {
      "Python parent id": 3696, "Python id": 3716, "finished": true, "Ev Idx": 10546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564849645.588, "dur": 3.467,
    "args": {
      "Python parent id": 3647, "Python id": 3717, "finished": true, "Ev Idx": 10547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564849649.742, "dur": 0.273,
    "args": {
      "Python parent id": 3647, "Python id": 3718, "finished": true, "Ev Idx": 10548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849651.048, "dur": 0.896,
    "args": {
      "Python parent id": 3647, "Python id": 3719, "finished": true, "Ev Idx": 10549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_157", "pid": 537, "tid": 537,
    "ts": 7429564849656.492, "dur": 44.032,
    "args": {
      "Python parent id": 3647, "Python module id": 157, "Python id": 3720, "finished": true, "Ev Idx": 10550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564849657.575, "dur": 42.550,
    "args": {
      "Python parent id": 3720, "Python id": 3721, "finished": true, "Ev Idx": 10551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564849658.045, "dur": 1.059,
    "args": {
      "Python parent id": 3721, "Python id": 3722, "finished": true, "Ev Idx": 10552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564849660.509, "dur": 39.405,
    "args": {
      "Python parent id": 3721, "Python id": 3723, "finished": true, "Ev Idx": 10553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849660.913, "dur": 0.180,
    "args": {
      "Python parent id": 3723, "Python id": 3724, "finished": true, "Ev Idx": 10554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849661.370, "dur": 0.304,
    "args": {
      "Python parent id": 3723, "Python id": 3725, "finished": true, "Ev Idx": 10555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564849661.846, "dur": 37.930,
    "args": {
      "Python parent id": 3723, "Python id": 3726, "finished": true, "Ev Idx": 10556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849714.759, "dur": 0.668,
    "args": {
      "Python parent id": 3631, "Python id": 3727, "finished": true, "Ev Idx": 10557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_45", "pid": 537, "tid": 537,
    "ts": 7429564849717.889, "dur": 83.180,
    "args": {
      "Python parent id": 3631, "Python module id": 45, "Python id": 3728, "finished": true, "Ev Idx": 10558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564849718.526, "dur": 82.245,
    "args": {
      "Python parent id": 3728, "Python id": 3729, "finished": true, "Ev Idx": 10559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564849718.825, "dur": 0.405,
    "args": {
      "Python parent id": 3729, "Python id": 3730, "finished": true, "Ev Idx": 10560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564849720.028, "dur": 79.877,
    "args": {
      "Python parent id": 3729, "Python id": 3731, "finished": true, "Ev Idx": 10561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564849720.582, "dur": 1.679,
    "args": {
      "Python parent id": 3731, "Python id": 3732, "finished": true, "Ev Idx": 10562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564849722.565, "dur": 12.518,
    "args": {
      "Python parent id": 3731, "Python id": 3733, "finished": true, "Ev Idx": 10563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564849735.490, "dur": 14.481,
    "args": {
      "Python parent id": 3731, "Python id": 3734, "finished": true, "Ev Idx": 10564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564849763.369, "dur": 11.491,
    "args": {
      "Python parent id": 3731, "Python id": 3735, "finished": true, "Ev Idx": 10565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849786.304, "dur": 0.249,
    "args": {
      "Python parent id": 3731, "Python id": 3736, "finished": true, "Ev Idx": 10566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564849786.955, "dur": 1.350,
    "args": {
      "Python parent id": 3731, "Python id": 3737, "finished": true, "Ev Idx": 10567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849801.843, "dur": 0.389,
    "args": {
      "Python parent id": 3631, "Python id": 3738, "finished": true, "Ev Idx": 10568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_22", "pid": 537, "tid": 537,
    "ts": 7429564849803.682, "dur": 298.416,
    "args": {
      "Python parent id": 3631, "Python module id": 22, "Python id": 3739, "finished": true, "Ev Idx": 10569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564849804.231, "dur": 297.519,
    "args": {
      "Python parent id": 3739, "Python id": 3740, "finished": true, "Ev Idx": 10570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564849804.416, "dur": 0.348,
    "args": {
      "Python parent id": 3740, "Python id": 3741, "finished": true, "Ev Idx": 10571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564849805.630, "dur": 295.949,
    "args": {
      "Python parent id": 3740, "Python id": 3742, "finished": true, "Ev Idx": 10572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849805.914, "dur": 0.322,
    "args": {
      "Python parent id": 3742, "Python id": 3743, "finished": true, "Ev Idx": 10573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849806.601, "dur": 0.331,
    "args": {
      "Python parent id": 3742, "Python id": 3744, "finished": true, "Ev Idx": 10574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849807.155, "dur": 0.260,
    "args": {
      "Python parent id": 3742, "Python id": 3745, "finished": true, "Ev Idx": 10575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_158", "pid": 537, "tid": 537,
    "ts": 7429564849809.331, "dur": 53.919,
    "args": {
      "Python parent id": 3742, "Python module id": 158, "Python id": 3746, "finished": true, "Ev Idx": 10576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564849809.952, "dur": 52.906,
    "args": {
      "Python parent id": 3746, "Python id": 3747, "finished": true, "Ev Idx": 10577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564849810.121, "dur": 0.164,
    "args": {
      "Python parent id": 3747, "Python id": 3748, "finished": true, "Ev Idx": 10578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564849811.036, "dur": 51.709,
    "args": {
      "Python parent id": 3747, "Python id": 3749, "finished": true, "Ev Idx": 10579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849811.337, "dur": 0.118,
    "args": {
      "Python parent id": 3749, "Python id": 3750, "finished": true, "Ev Idx": 10580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849811.681, "dur": 0.110,
    "args": {
      "Python parent id": 3749, "Python id": 3751, "finished": true, "Ev Idx": 10581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564849811.944, "dur": 50.697,
    "args": {
      "Python parent id": 3749, "Python id": 3752, "finished": true, "Ev Idx": 10582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_22", "pid": 537, "tid": 537,
    "ts": 7429564849865.097, "dur": 67.305,
    "args": {
      "Python parent id": 3742, "Python module id": 22, "Python id": 3753, "finished": true, "Ev Idx": 10583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564849865.575, "dur": 66.503,
    "args": {
      "Python parent id": 3753, "Python id": 3754, "finished": true, "Ev Idx": 10584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564849865.802, "dur": 0.291,
    "args": {
      "Python parent id": 3754, "Python id": 3755, "finished": true, "Ev Idx": 10585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564849867.123, "dur": 64.652,
    "args": {
      "Python parent id": 3754, "Python id": 3756, "finished": true, "Ev Idx": 10586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564849867.955, "dur": 62.179,
    "args": {
      "Python parent id": 3756, "Python id": 3757, "finished": true, "Ev Idx": 10587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564849868.105, "dur": 0.165,
    "args": {
      "Python parent id": 3757, "Python id": 3758, "finished": true, "Ev Idx": 10588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564849869.215, "dur": 60.748,
    "args": {
      "Python parent id": 3757, "Python id": 3759, "finished": true, "Ev Idx": 10589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849933.439, "dur": 0.459,
    "args": {
      "Python parent id": 3742, "Python id": 3760, "finished": true, "Ev Idx": 10590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_159", "pid": 537, "tid": 537,
    "ts": 7429564849936.218, "dur": 51.545,
    "args": {
      "Python parent id": 3742, "Python module id": 159, "Python id": 3761, "finished": true, "Ev Idx": 10591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564849936.631, "dur": 50.879,
    "args": {
      "Python parent id": 3761, "Python id": 3762, "finished": true, "Ev Idx": 10592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564849936.844, "dur": 0.316,
    "args": {
      "Python parent id": 3762, "Python id": 3763, "finished": true, "Ev Idx": 10593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564849937.856, "dur": 49.553,
    "args": {
      "Python parent id": 3762, "Python id": 3764, "finished": true, "Ev Idx": 10594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849938.189, "dur": 0.274,
    "args": {
      "Python parent id": 3764, "Python id": 3765, "finished": true, "Ev Idx": 10595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564849938.664, "dur": 0.140,
    "args": {
      "Python parent id": 3764, "Python id": 3766, "finished": true, "Ev Idx": 10596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564849938.927, "dur": 48.389,
    "args": {
      "Python parent id": 3764, "Python id": 3767, "finished": true, "Ev Idx": 10597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_160", "pid": 537, "tid": 537,
    "ts": 7429564850034.395, "dur": 66.664,
    "args": {
      "Python parent id": 3742, "Python module id": 160, "Python id": 3768, "finished": true, "Ev Idx": 10598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850035.070, "dur": 65.686,
    "args": {
      "Python parent id": 3768, "Python id": 3769, "finished": true, "Ev Idx": 10599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850035.417, "dur": 0.368,
    "args": {
      "Python parent id": 3769, "Python id": 3770, "finished": true, "Ev Idx": 10600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564850036.401, "dur": 64.083,
    "args": {
      "Python parent id": 3769, "Python id": 3771, "finished": true, "Ev Idx": 10601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850036.759, "dur": 0.194,
    "args": {
      "Python parent id": 3771, "Python id": 3772, "finished": true, "Ev Idx": 10602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850037.157, "dur": 0.143,
    "args": {
      "Python parent id": 3771, "Python id": 3773, "finished": true, "Ev Idx": 10603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564850037.403, "dur": 62.995,
    "args": {
      "Python parent id": 3771, "Python id": 3774, "finished": true, "Ev Idx": 10604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_23", "pid": 537, "tid": 537,
    "ts": 7429564850125.999, "dur": 1199.540,
    "args": {
      "Python parent id": 64, "Python module id": 23, "Python id": 3775, "finished": true, "Ev Idx": 10605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850127.401, "dur": 1197.212,
    "args": {
      "Python parent id": 3775, "Python id": 3776, "finished": true, "Ev Idx": 10606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850127.637, "dur": 0.456,
    "args": {
      "Python parent id": 3776, "Python id": 3777, "finished": true, "Ev Idx": 10607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564850129.777, "dur": 1193.800,
    "args": {
      "Python parent id": 3776, "Python id": 3778, "finished": true, "Ev Idx": 10608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850130.253, "dur": 0.387,
    "args": {
      "Python parent id": 3778, "Python id": 3779, "finished": true, "Ev Idx": 10609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_46", "pid": 537, "tid": 537,
    "ts": 7429564850132.936, "dur": 83.280,
    "args": {
      "Python parent id": 3778, "Python module id": 46, "Python id": 3780, "finished": true, "Ev Idx": 10610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850133.597, "dur": 82.284,
    "args": {
      "Python parent id": 3780, "Python id": 3781, "finished": true, "Ev Idx": 10611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850133.747, "dur": 0.144,
    "args": {
      "Python parent id": 3781, "Python id": 3782, "finished": true, "Ev Idx": 10612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564850134.866, "dur": 80.011,
    "args": {
      "Python parent id": 3781, "Python id": 3783, "finished": true, "Ev Idx": 10613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564850135.434, "dur": 2.021,
    "args": {
      "Python parent id": 3783, "Python id": 3784, "finished": true, "Ev Idx": 10614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564850137.738, "dur": 14.163,
    "args": {
      "Python parent id": 3783, "Python id": 3785, "finished": true, "Ev Idx": 10615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564850152.336, "dur": 15.382,
    "args": {
      "Python parent id": 3783, "Python id": 3786, "finished": true, "Ev Idx": 10616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564850182.491, "dur": 11.070,
    "args": {
      "Python parent id": 3783, "Python id": 3787, "finished": true, "Ev Idx": 10617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850203.712, "dur": 0.273,
    "args": {
      "Python parent id": 3783, "Python id": 3788, "finished": true, "Ev Idx": 10618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564850204.454, "dur": 1.266,
    "args": {
      "Python parent id": 3783, "Python id": 3789, "finished": true, "Ev Idx": 10619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850216.845, "dur": 0.323,
    "args": {
      "Python parent id": 3778, "Python id": 3790, "finished": true, "Ev Idx": 10620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_23", "pid": 537, "tid": 537,
    "ts": 7429564850219.684, "dur": 662.664,
    "args": {
      "Python parent id": 3778, "Python module id": 23, "Python id": 3791, "finished": true, "Ev Idx": 10621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850221.539, "dur": 660.195,
    "args": {
      "Python parent id": 3791, "Python id": 3792, "finished": true, "Ev Idx": 10622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850221.714, "dur": 0.333,
    "args": {
      "Python parent id": 3792, "Python id": 3793, "finished": true, "Ev Idx": 10623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564850223.749, "dur": 657.088,
    "args": {
      "Python parent id": 3792, "Python id": 3794, "finished": true, "Ev Idx": 10624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850227.627, "dur": 0.337,
    "args": {
      "Python parent id": 3794, "Python id": 3795, "finished": true, "Ev Idx": 10625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_161", "pid": 537, "tid": 537,
    "ts": 7429564850230.029, "dur": 45.341,
    "args": {
      "Python parent id": 3794, "Python module id": 161, "Python id": 3796, "finished": true, "Ev Idx": 10626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850230.594, "dur": 44.519,
    "args": {
      "Python parent id": 3796, "Python id": 3797, "finished": true, "Ev Idx": 10627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850230.749, "dur": 0.182,
    "args": {
      "Python parent id": 3797, "Python id": 3798, "finished": true, "Ev Idx": 10628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564850231.680, "dur": 43.333,
    "args": {
      "Python parent id": 3797, "Python id": 3799, "finished": true, "Ev Idx": 10629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850231.969, "dur": 0.118,
    "args": {
      "Python parent id": 3799, "Python id": 3800, "finished": true, "Ev Idx": 10630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850232.277, "dur": 0.138,
    "args": {
      "Python parent id": 3799, "Python id": 3801, "finished": true, "Ev Idx": 10631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564850232.538, "dur": 42.403,
    "args": {
      "Python parent id": 3799, "Python id": 3802, "finished": true, "Ev Idx": 10632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564850276.085, "dur": 3.456,
    "args": {
      "Python parent id": 3794, "Python id": 3803, "finished": true, "Ev Idx": 10633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564850279.991, "dur": 4.166,
    "args": {
      "Python parent id": 3794, "Python id": 3804, "finished": true, "Ev Idx": 10634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850284.948, "dur": 0.477,
    "args": {
      "Python parent id": 3794, "Python id": 3805, "finished": true, "Ev Idx": 10635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_162", "pid": 537, "tid": 537,
    "ts": 7429564850287.631, "dur": 46.109,
    "args": {
      "Python parent id": 3794, "Python module id": 162, "Python id": 3806, "finished": true, "Ev Idx": 10636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850288.179, "dur": 45.384,
    "args": {
      "Python parent id": 3806, "Python id": 3807, "finished": true, "Ev Idx": 10637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850288.365, "dur": 0.313,
    "args": {
      "Python parent id": 3807, "Python id": 3808, "finished": true, "Ev Idx": 10638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564850289.264, "dur": 44.184,
    "args": {
      "Python parent id": 3807, "Python id": 3809, "finished": true, "Ev Idx": 10639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850289.496, "dur": 0.164,
    "args": {
      "Python parent id": 3809, "Python id": 3810, "finished": true, "Ev Idx": 10640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850289.838, "dur": 0.149,
    "args": {
      "Python parent id": 3809, "Python id": 3811, "finished": true, "Ev Idx": 10641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564850290.098, "dur": 43.301,
    "args": {
      "Python parent id": 3809, "Python id": 3812, "finished": true, "Ev Idx": 10642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564850334.290, "dur": 2.593,
    "args": {
      "Python parent id": 3794, "Python id": 3813, "finished": true, "Ev Idx": 10643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564850337.345, "dur": 2.530,
    "args": {
      "Python parent id": 3794, "Python id": 3814, "finished": true, "Ev Idx": 10644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850340.757, "dur": 0.468,
    "args": {
      "Python parent id": 3794, "Python id": 3815, "finished": true, "Ev Idx": 10645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_163", "pid": 537, "tid": 537,
    "ts": 7429564850343.357, "dur": 41.218,
    "args": {
      "Python parent id": 3794, "Python module id": 163, "Python id": 3816, "finished": true, "Ev Idx": 10646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850343.916, "dur": 40.453,
    "args": {
      "Python parent id": 3816, "Python id": 3817, "finished": true, "Ev Idx": 10647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850344.106, "dur": 0.291,
    "args": {
      "Python parent id": 3817, "Python id": 3818, "finished": true, "Ev Idx": 10648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564850345.090, "dur": 39.163,
    "args": {
      "Python parent id": 3817, "Python id": 3819, "finished": true, "Ev Idx": 10649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850345.311, "dur": 0.129,
    "args": {
      "Python parent id": 3819, "Python id": 3820, "finished": true, "Ev Idx": 10650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850345.629, "dur": 0.118,
    "args": {
      "Python parent id": 3819, "Python id": 3821, "finished": true, "Ev Idx": 10651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564850345.853, "dur": 38.352,
    "args": {
      "Python parent id": 3819, "Python id": 3822, "finished": true, "Ev Idx": 10652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564850385.075, "dur": 2.323,
    "args": {
      "Python parent id": 3794, "Python id": 3823, "finished": true, "Ev Idx": 10653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564850387.901, "dur": 2.396,
    "args": {
      "Python parent id": 3794, "Python id": 3824, "finished": true, "Ev Idx": 10654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564850391.394, "dur": 147.563,
    "args": {
      "Python parent id": 3794, "Python id": 3825, "finished": true, "Ev Idx": 10655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564850391.631, "dur": 3.675,
    "args": {
      "Python parent id": 3825, "Python id": 3826, "finished": true, "Ev Idx": 10656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564850395.596, "dur": 1.782,
    "args": {
      "Python parent id": 3825, "Python id": 3827, "finished": true, "Ev Idx": 10657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564850408.685, "dur": 50.994,
    "args": {
      "Python parent id": 3825, "Python id": 3828, "finished": true, "Ev Idx": 10658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564850436.525, "dur": 22.082,
    "args": {
      "Python parent id": 3828, "Python id": 3829, "finished": true, "Ev Idx": 10659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564850489.876, "dur": 29.586,
    "args": {
      "Python parent id": 3825, "Python id": 3830, "finished": true, "Ev Idx": 10660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564850507.596, "dur": 11.052,
    "args": {
      "Python parent id": 3830, "Python id": 3831, "finished": true, "Ev Idx": 10661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564850543.177, "dur": 4.101,
    "args": {
      "Python parent id": 3794, "Python id": 3832, "finished": true, "Ev Idx": 10662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564850544.051, "dur": 0.341,
    "args": {
      "Python parent id": 3832, "Python id": 3833, "finished": true, "Ev Idx": 10663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564850544.795, "dur": 0.092,
    "args": {
      "Python parent id": 3832, "Python id": 3834, "finished": true, "Ev Idx": 10664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564850546.273, "dur": 0.240,
    "args": {
      "Python parent id": 3832, "Python id": 3835, "finished": true, "Ev Idx": 10665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564850546.871, "dur": 0.133,
    "args": {
      "Python parent id": 3832, "Python id": 3836, "finished": true, "Ev Idx": 10666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564850548.667, "dur": 6.952,
    "args": {
      "Python parent id": 3794, "Python id": 3837, "finished": true, "Ev Idx": 10667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564850551.950, "dur": 3.509,
    "args": {
      "Python parent id": 3837, "Python id": 3838, "finished": true, "Ev Idx": 10668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564850552.186, "dur": 1.630,
    "args": {
      "Python parent id": 3838, "Python id": 3839, "finished": true, "Ev Idx": 10669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564850552.758, "dur": 0.866,
    "args": {
      "Python parent id": 3839, "Python id": 3840, "finished": true, "Ev Idx": 10670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564850554.077, "dur": 0.523,
    "args": {
      "Python parent id": 3838, "Python id": 3841, "finished": true, "Ev Idx": 10671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564850554.939, "dur": 0.447,
    "args": {
      "Python parent id": 3838, "Python id": 3842, "finished": true, "Ev Idx": 10672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564850558.162, "dur": 264.680,
    "args": {
      "Python parent id": 3794, "Python id": 3843, "finished": true, "Ev Idx": 10673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564850558.832, "dur": 48.692,
    "args": {
      "Python parent id": 3843, "Python id": 3844, "finished": true, "Ev Idx": 10674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564850577.073, "dur": 6.181,
    "args": {
      "Python parent id": 3844, "Python id": 3845, "finished": true, "Ev Idx": 10675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564850584.086, "dur": 23.307,
    "args": {
      "Python parent id": 3844, "Python id": 3846, "finished": true, "Ev Idx": 10676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564850608.391, "dur": 33.877,
    "args": {
      "Python parent id": 3843, "Python id": 3847, "finished": true, "Ev Idx": 10677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564850625.887, "dur": 2.647,
    "args": {
      "Python parent id": 3847, "Python id": 3848, "finished": true, "Ev Idx": 10678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564850629.005, "dur": 13.157,
    "args": {
      "Python parent id": 3847, "Python id": 3849, "finished": true, "Ev Idx": 10679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564850643.143, "dur": 3.244,
    "args": {
      "Python parent id": 3843, "Python id": 3850, "finished": true, "Ev Idx": 10680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564850646.709, "dur": 46.834,
    "args": {
      "Python parent id": 3843, "Python id": 3851, "finished": true, "Ev Idx": 10681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564850727.891, "dur": 18.167,
    "args": {
      "Python parent id": 3843, "Python id": 3852, "finished": true, "Ev Idx": 10682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564850728.198, "dur": 0.241,
    "args": {
      "Python parent id": 3852, "Python id": 3853, "finished": true, "Ev Idx": 10683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564850729.152, "dur": 16.768,
    "args": {
      "Python parent id": 3852, "Python id": 3854, "finished": true, "Ev Idx": 10684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564850746.989, "dur": 1.681,
    "args": {
      "Python parent id": 3843, "Python id": 3855, "finished": true, "Ev Idx": 10685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564850750.238, "dur": 10.441,
    "args": {
      "Python parent id": 3843, "Python id": 3856, "finished": true, "Ev Idx": 10686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564850750.359, "dur": 0.118,
    "args": {
      "Python parent id": 3856, "Python id": 3857, "finished": true, "Ev Idx": 10687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850756.087, "dur": 1.077,
    "args": {
      "Python parent id": 3856, "Python id": 3858, "finished": true, "Ev Idx": 10688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564850756.589, "dur": 0.495,
    "args": {
      "Python parent id": 3858, "Python id": 3859, "finished": true, "Ev Idx": 10689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564850757.524, "dur": 3.055,
    "args": {
      "Python parent id": 3856, "Python id": 3860, "finished": true, "Ev Idx": 10690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564850761.414, "dur": 43.307,
    "args": {
      "Python parent id": 3843, "Python id": 3861, "finished": true, "Ev Idx": 10691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564850805.346, "dur": 2.987,
    "args": {
      "Python parent id": 3843, "Python id": 3862, "finished": true, "Ev Idx": 10692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564850809.089, "dur": 12.709,
    "args": {
      "Python parent id": 3843, "Python id": 3863, "finished": true, "Ev Idx": 10693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564850825.590, "dur": 3.278,
    "args": {
      "Python parent id": 3794, "Python id": 3864, "finished": true, "Ev Idx": 10694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564850829.560, "dur": 0.283,
    "args": {
      "Python parent id": 3794, "Python id": 3865, "finished": true, "Ev Idx": 10695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850830.875, "dur": 0.885,
    "args": {
      "Python parent id": 3794, "Python id": 3866, "finished": true, "Ev Idx": 10696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_164", "pid": 537, "tid": 537,
    "ts": 7429564850836.136, "dur": 44.050,
    "args": {
      "Python parent id": 3794, "Python module id": 164, "Python id": 3867, "finished": true, "Ev Idx": 10697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850837.242, "dur": 42.492,
    "args": {
      "Python parent id": 3867, "Python id": 3868, "finished": true, "Ev Idx": 10698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850837.705, "dur": 1.075,
    "args": {
      "Python parent id": 3868, "Python id": 3869, "finished": true, "Ev Idx": 10699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564850840.079, "dur": 39.479,
    "args": {
      "Python parent id": 3868, "Python id": 3870, "finished": true, "Ev Idx": 10700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850840.511, "dur": 0.246,
    "args": {
      "Python parent id": 3870, "Python id": 3871, "finished": true, "Ev Idx": 10701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850841.050, "dur": 0.166,
    "args": {
      "Python parent id": 3870, "Python id": 3872, "finished": true, "Ev Idx": 10702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564850841.416, "dur": 38.016,
    "args": {
      "Python parent id": 3870, "Python id": 3873, "finished": true, "Ev Idx": 10703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850893.808, "dur": 0.504,
    "args": {
      "Python parent id": 3778, "Python id": 3874, "finished": true, "Ev Idx": 10704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_47", "pid": 537, "tid": 537,
    "ts": 7429564850897.029, "dur": 78.298,
    "args": {
      "Python parent id": 3778, "Python module id": 47, "Python id": 3875, "finished": true, "Ev Idx": 10705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850897.683, "dur": 77.164,
    "args": {
      "Python parent id": 3875, "Python id": 3876, "finished": true, "Ev Idx": 10706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850897.898, "dur": 0.473,
    "args": {
      "Python parent id": 3876, "Python id": 3877, "finished": true, "Ev Idx": 10707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564850899.325, "dur": 74.610,
    "args": {
      "Python parent id": 3876, "Python id": 3878, "finished": true, "Ev Idx": 10708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564850899.795, "dur": 1.431,
    "args": {
      "Python parent id": 3878, "Python id": 3879, "finished": true, "Ev Idx": 10709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564850901.518, "dur": 13.327,
    "args": {
      "Python parent id": 3878, "Python id": 3880, "finished": true, "Ev Idx": 10710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564850915.248, "dur": 14.276,
    "args": {
      "Python parent id": 3878, "Python id": 3881, "finished": true, "Ev Idx": 10711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564850941.579, "dur": 10.738,
    "args": {
      "Python parent id": 3878, "Python id": 3882, "finished": true, "Ev Idx": 10712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850962.356, "dur": 0.192,
    "args": {
      "Python parent id": 3878, "Python id": 3883, "finished": true, "Ev Idx": 10713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564850962.923, "dur": 1.404,
    "args": {
      "Python parent id": 3878, "Python id": 3884, "finished": true, "Ev Idx": 10714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850976.097, "dur": 0.556,
    "args": {
      "Python parent id": 3778, "Python id": 3885, "finished": true, "Ev Idx": 10715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_23", "pid": 537, "tid": 537,
    "ts": 7429564850978.409, "dur": 333.782,
    "args": {
      "Python parent id": 3778, "Python module id": 23, "Python id": 3886, "finished": true, "Ev Idx": 10716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850979.013, "dur": 332.916,
    "args": {
      "Python parent id": 3886, "Python id": 3887, "finished": true, "Ev Idx": 10717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850979.228, "dur": 0.349,
    "args": {
      "Python parent id": 3887, "Python id": 3888, "finished": true, "Ev Idx": 10718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564850980.594, "dur": 331.159,
    "args": {
      "Python parent id": 3887, "Python id": 3889, "finished": true, "Ev Idx": 10719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850980.893, "dur": 0.342,
    "args": {
      "Python parent id": 3889, "Python id": 3890, "finished": true, "Ev Idx": 10720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850981.660, "dur": 0.305,
    "args": {
      "Python parent id": 3889, "Python id": 3891, "finished": true, "Ev Idx": 10721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850982.280, "dur": 0.324,
    "args": {
      "Python parent id": 3889, "Python id": 3892, "finished": true, "Ev Idx": 10722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_165", "pid": 537, "tid": 537,
    "ts": 7429564850984.274, "dur": 124.423,
    "args": {
      "Python parent id": 3889, "Python module id": 165, "Python id": 3893, "finished": true, "Ev Idx": 10723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564850984.977, "dur": 123.268,
    "args": {
      "Python parent id": 3893, "Python id": 3894, "finished": true, "Ev Idx": 10724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564850985.118, "dur": 0.174,
    "args": {
      "Python parent id": 3894, "Python id": 3895, "finished": true, "Ev Idx": 10725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564850986.147, "dur": 121.827,
    "args": {
      "Python parent id": 3894, "Python id": 3896, "finished": true, "Ev Idx": 10726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850986.505, "dur": 0.162,
    "args": {
      "Python parent id": 3896, "Python id": 3897, "finished": true, "Ev Idx": 10727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564850986.937, "dur": 0.195,
    "args": {
      "Python parent id": 3896, "Python id": 3898, "finished": true, "Ev Idx": 10728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564850987.258, "dur": 120.533,
    "args": {
      "Python parent id": 3896, "Python id": 3899, "finished": true, "Ev Idx": 10729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_23", "pid": 537, "tid": 537,
    "ts": 7429564851110.950, "dur": 19.160,
    "args": {
      "Python parent id": 3889, "Python module id": 23, "Python id": 3900, "finished": true, "Ev Idx": 10730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564851111.550, "dur": 18.248,
    "args": {
      "Python parent id": 3900, "Python id": 3901, "finished": true, "Ev Idx": 10731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564851111.793, "dur": 0.463,
    "args": {
      "Python parent id": 3901, "Python id": 3902, "finished": true, "Ev Idx": 10732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564851113.662, "dur": 16.062,
    "args": {
      "Python parent id": 3901, "Python id": 3903, "finished": true, "Ev Idx": 10733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564851114.611, "dur": 14.964,
    "args": {
      "Python parent id": 3903, "Python id": 3904, "finished": true, "Ev Idx": 10734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564851114.786, "dur": 0.258,
    "args": {
      "Python parent id": 3904, "Python id": 3905, "finished": true, "Ev Idx": 10735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564851116.017, "dur": 13.474,
    "args": {
      "Python parent id": 3904, "Python id": 3906, "finished": true, "Ev Idx": 10736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851131.047, "dur": 0.491,
    "args": {
      "Python parent id": 3889, "Python id": 3907, "finished": true, "Ev Idx": 10737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_166", "pid": 537, "tid": 537,
    "ts": 7429564851133.791, "dur": 116.928,
    "args": {
      "Python parent id": 3889, "Python module id": 166, "Python id": 3908, "finished": true, "Ev Idx": 10738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564851134.366, "dur": 116.112,
    "args": {
      "Python parent id": 3908, "Python id": 3909, "finished": true, "Ev Idx": 10739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564851134.564, "dur": 0.223,
    "args": {
      "Python parent id": 3909, "Python id": 3910, "finished": true, "Ev Idx": 10740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564851135.435, "dur": 114.838,
    "args": {
      "Python parent id": 3909, "Python id": 3911, "finished": true, "Ev Idx": 10741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851135.693, "dur": 0.137,
    "args": {
      "Python parent id": 3911, "Python id": 3912, "finished": true, "Ev Idx": 10742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851136.036, "dur": 0.124,
    "args": {
      "Python parent id": 3911, "Python id": 3913, "finished": true, "Ev Idx": 10743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564851136.334, "dur": 113.815,
    "args": {
      "Python parent id": 3911, "Python id": 3914, "finished": true, "Ev Idx": 10744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_167", "pid": 537, "tid": 537,
    "ts": 7429564851264.575, "dur": 46.912,
    "args": {
      "Python parent id": 3889, "Python module id": 167, "Python id": 3915, "finished": true, "Ev Idx": 10745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564851265.295, "dur": 45.921,
    "args": {
      "Python parent id": 3915, "Python id": 3916, "finished": true, "Ev Idx": 10746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564851265.559, "dur": 0.393,
    "args": {
      "Python parent id": 3916, "Python id": 3917, "finished": true, "Ev Idx": 10747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564851266.765, "dur": 44.342,
    "args": {
      "Python parent id": 3916, "Python id": 3918, "finished": true, "Ev Idx": 10748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851267.074, "dur": 0.187,
    "args": {
      "Python parent id": 3918, "Python id": 3919, "finished": true, "Ev Idx": 10749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851267.503, "dur": 0.142,
    "args": {
      "Python parent id": 3918, "Python id": 3920, "finished": true, "Ev Idx": 10750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564851267.759, "dur": 43.283,
    "args": {
      "Python parent id": 3918, "Python id": 3921, "finished": true, "Ev Idx": 10751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_24", "pid": 537, "tid": 537,
    "ts": 7429564851332.471, "dur": 1173.321,
    "args": {
      "Python parent id": 64, "Python module id": 24, "Python id": 3922, "finished": true, "Ev Idx": 10752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564851334.011, "dur": 1170.962,
    "args": {
      "Python parent id": 3922, "Python id": 3923, "finished": true, "Ev Idx": 10753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564851334.238, "dur": 0.325,
    "args": {
      "Python parent id": 3923, "Python id": 3924, "finished": true, "Ev Idx": 10754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564851336.368, "dur": 1167.851,
    "args": {
      "Python parent id": 3923, "Python id": 3925, "finished": true, "Ev Idx": 10755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851336.863, "dur": 0.655,
    "args": {
      "Python parent id": 3925, "Python id": 3926, "finished": true, "Ev Idx": 10756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_48", "pid": 537, "tid": 537,
    "ts": 7429564851339.352, "dur": 80.741,
    "args": {
      "Python parent id": 3925, "Python module id": 48, "Python id": 3927, "finished": true, "Ev Idx": 10757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564851339.987, "dur": 79.762,
    "args": {
      "Python parent id": 3927, "Python id": 3928, "finished": true, "Ev Idx": 10758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564851340.140, "dur": 0.189,
    "args": {
      "Python parent id": 3928, "Python id": 3929, "finished": true, "Ev Idx": 10759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564851341.094, "dur": 77.980,
    "args": {
      "Python parent id": 3928, "Python id": 3930, "finished": true, "Ev Idx": 10760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564851341.656, "dur": 1.928,
    "args": {
      "Python parent id": 3930, "Python id": 3931, "finished": true, "Ev Idx": 10761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564851343.903, "dur": 11.559,
    "args": {
      "Python parent id": 3930, "Python id": 3932, "finished": true, "Ev Idx": 10762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564851355.847, "dur": 15.807,
    "args": {
      "Python parent id": 3930, "Python id": 3933, "finished": true, "Ev Idx": 10763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564851383.251, "dur": 12.909,
    "args": {
      "Python parent id": 3930, "Python id": 3934, "finished": true, "Ev Idx": 10764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851406.990, "dur": 0.211,
    "args": {
      "Python parent id": 3930, "Python id": 3935, "finished": true, "Ev Idx": 10765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564851407.583, "dur": 1.207,
    "args": {
      "Python parent id": 3930, "Python id": 3936, "finished": true, "Ev Idx": 10766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851420.818, "dur": 0.364,
    "args": {
      "Python parent id": 3925, "Python id": 3937, "finished": true, "Ev Idx": 10767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_24", "pid": 537, "tid": 537,
    "ts": 7429564851423.838, "dur": 691.623,
    "args": {
      "Python parent id": 3925, "Python module id": 24, "Python id": 3938, "finished": true, "Ev Idx": 10768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564851425.340, "dur": 689.299,
    "args": {
      "Python parent id": 3938, "Python id": 3939, "finished": true, "Ev Idx": 10769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564851425.596, "dur": 0.332,
    "args": {
      "Python parent id": 3939, "Python id": 3940, "finished": true, "Ev Idx": 10770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564851427.577, "dur": 686.294,
    "args": {
      "Python parent id": 3939, "Python id": 3941, "finished": true, "Ev Idx": 10771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851431.022, "dur": 0.368,
    "args": {
      "Python parent id": 3941, "Python id": 3942, "finished": true, "Ev Idx": 10772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_168", "pid": 537, "tid": 537,
    "ts": 7429564851433.447, "dur": 41.958,
    "args": {
      "Python parent id": 3941, "Python module id": 168, "Python id": 3943, "finished": true, "Ev Idx": 10773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564851434.104, "dur": 41.044,
    "args": {
      "Python parent id": 3943, "Python id": 3944, "finished": true, "Ev Idx": 10774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564851434.283, "dur": 0.201,
    "args": {
      "Python parent id": 3944, "Python id": 3945, "finished": true, "Ev Idx": 10775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564851435.312, "dur": 39.729,
    "args": {
      "Python parent id": 3944, "Python id": 3946, "finished": true, "Ev Idx": 10776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851435.654, "dur": 0.147,
    "args": {
      "Python parent id": 3946, "Python id": 3947, "finished": true, "Ev Idx": 10777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851436.056, "dur": 0.104,
    "args": {
      "Python parent id": 3946, "Python id": 3948, "finished": true, "Ev Idx": 10778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564851436.284, "dur": 38.701,
    "args": {
      "Python parent id": 3946, "Python id": 3949, "finished": true, "Ev Idx": 10779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564851476.144, "dur": 3.237,
    "args": {
      "Python parent id": 3941, "Python id": 3950, "finished": true, "Ev Idx": 10780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564851479.845, "dur": 3.845,
    "args": {
      "Python parent id": 3941, "Python id": 3951, "finished": true, "Ev Idx": 10781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851484.403, "dur": 0.346,
    "args": {
      "Python parent id": 3941, "Python id": 3952, "finished": true, "Ev Idx": 10782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_169", "pid": 537, "tid": 537,
    "ts": 7429564851486.842, "dur": 46.672,
    "args": {
      "Python parent id": 3941, "Python module id": 169, "Python id": 3953, "finished": true, "Ev Idx": 10783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564851487.363, "dur": 45.897,
    "args": {
      "Python parent id": 3953, "Python id": 3954, "finished": true, "Ev Idx": 10784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564851487.602, "dur": 0.262,
    "args": {
      "Python parent id": 3954, "Python id": 3955, "finished": true, "Ev Idx": 10785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564851488.503, "dur": 44.616,
    "args": {
      "Python parent id": 3954, "Python id": 3956, "finished": true, "Ev Idx": 10786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851488.823, "dur": 0.338,
    "args": {
      "Python parent id": 3956, "Python id": 3957, "finished": true, "Ev Idx": 10787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851489.535, "dur": 0.113,
    "args": {
      "Python parent id": 3956, "Python id": 3958, "finished": true, "Ev Idx": 10788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564851489.781, "dur": 43.237,
    "args": {
      "Python parent id": 3956, "Python id": 3959, "finished": true, "Ev Idx": 10789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564851534.019, "dur": 2.444,
    "args": {
      "Python parent id": 3941, "Python id": 3960, "finished": true, "Ev Idx": 10790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564851536.930, "dur": 2.741,
    "args": {
      "Python parent id": 3941, "Python id": 3961, "finished": true, "Ev Idx": 10791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851540.417, "dur": 0.449,
    "args": {
      "Python parent id": 3941, "Python id": 3962, "finished": true, "Ev Idx": 10792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_170", "pid": 537, "tid": 537,
    "ts": 7429564851543.149, "dur": 41.567,
    "args": {
      "Python parent id": 3941, "Python module id": 170, "Python id": 3963, "finished": true, "Ev Idx": 10793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564851543.657, "dur": 40.753,
    "args": {
      "Python parent id": 3963, "Python id": 3964, "finished": true, "Ev Idx": 10794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564851543.835, "dur": 0.406,
    "args": {
      "Python parent id": 3964, "Python id": 3965, "finished": true, "Ev Idx": 10795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564851545.127, "dur": 39.172,
    "args": {
      "Python parent id": 3964, "Python id": 3966, "finished": true, "Ev Idx": 10796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851545.395, "dur": 0.175,
    "args": {
      "Python parent id": 3966, "Python id": 3967, "finished": true, "Ev Idx": 10797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851545.753, "dur": 0.149,
    "args": {
      "Python parent id": 3966, "Python id": 3968, "finished": true, "Ev Idx": 10798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564851546.025, "dur": 38.182,
    "args": {
      "Python parent id": 3966, "Python id": 3969, "finished": true, "Ev Idx": 10799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564851585.315, "dur": 2.466,
    "args": {
      "Python parent id": 3941, "Python id": 3970, "finished": true, "Ev Idx": 10800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564851588.261, "dur": 2.452,
    "args": {
      "Python parent id": 3941, "Python id": 3971, "finished": true, "Ev Idx": 10801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564851591.747, "dur": 192.792,
    "args": {
      "Python parent id": 3941, "Python id": 3972, "finished": true, "Ev Idx": 10802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564851592.081, "dur": 3.425,
    "args": {
      "Python parent id": 3972, "Python id": 3973, "finished": true, "Ev Idx": 10803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564851595.802, "dur": 1.805,
    "args": {
      "Python parent id": 3972, "Python id": 3974, "finished": true, "Ev Idx": 10804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564851649.386, "dur": 55.860,
    "args": {
      "Python parent id": 3972, "Python id": 3975, "finished": true, "Ev Idx": 10805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564851681.063, "dur": 22.885,
    "args": {
      "Python parent id": 3975, "Python id": 3976, "finished": true, "Ev Idx": 10806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564851737.992, "dur": 28.124,
    "args": {
      "Python parent id": 3972, "Python id": 3977, "finished": true, "Ev Idx": 10807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564851754.608, "dur": 10.595,
    "args": {
      "Python parent id": 3977, "Python id": 3978, "finished": true, "Ev Idx": 10808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564851789.350, "dur": 4.063,
    "args": {
      "Python parent id": 3941, "Python id": 3979, "finished": true, "Ev Idx": 10809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564851790.084, "dur": 0.355,
    "args": {
      "Python parent id": 3979, "Python id": 3980, "finished": true, "Ev Idx": 10810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564851790.767, "dur": 0.099,
    "args": {
      "Python parent id": 3979, "Python id": 3981, "finished": true, "Ev Idx": 10811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564851792.173, "dur": 0.463,
    "args": {
      "Python parent id": 3979, "Python id": 3982, "finished": true, "Ev Idx": 10812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564851793.052, "dur": 0.146,
    "args": {
      "Python parent id": 3979, "Python id": 3983, "finished": true, "Ev Idx": 10813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564851794.907, "dur": 6.897,
    "args": {
      "Python parent id": 3941, "Python id": 3984, "finished": true, "Ev Idx": 10814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564851798.012, "dur": 3.624,
    "args": {
      "Python parent id": 3984, "Python id": 3985, "finished": true, "Ev Idx": 10815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564851798.272, "dur": 1.473,
    "args": {
      "Python parent id": 3985, "Python id": 3986, "finished": true, "Ev Idx": 10816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564851798.740, "dur": 0.820,
    "args": {
      "Python parent id": 3986, "Python id": 3987, "finished": true, "Ev Idx": 10817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564851800.003, "dur": 0.519,
    "args": {
      "Python parent id": 3985, "Python id": 3988, "finished": true, "Ev Idx": 10818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564851800.943, "dur": 0.617,
    "args": {
      "Python parent id": 3985, "Python id": 3989, "finished": true, "Ev Idx": 10819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564851804.603, "dur": 252.478,
    "args": {
      "Python parent id": 3941, "Python id": 3990, "finished": true, "Ev Idx": 10820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564851805.307, "dur": 47.031,
    "args": {
      "Python parent id": 3990, "Python id": 3991, "finished": true, "Ev Idx": 10821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564851822.665, "dur": 5.661,
    "args": {
      "Python parent id": 3991, "Python id": 3992, "finished": true, "Ev Idx": 10822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564851829.162, "dur": 23.035,
    "args": {
      "Python parent id": 3991, "Python id": 3993, "finished": true, "Ev Idx": 10823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564851853.112, "dur": 31.455,
    "args": {
      "Python parent id": 3990, "Python id": 3994, "finished": true, "Ev Idx": 10824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564851868.343, "dur": 2.614,
    "args": {
      "Python parent id": 3994, "Python id": 3995, "finished": true, "Ev Idx": 10825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564851871.432, "dur": 13.071,
    "args": {
      "Python parent id": 3994, "Python id": 3996, "finished": true, "Ev Idx": 10826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564851885.417, "dur": 4.067,
    "args": {
      "Python parent id": 3990, "Python id": 3997, "finished": true, "Ev Idx": 10827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564851889.777, "dur": 46.301,
    "args": {
      "Python parent id": 3990, "Python id": 3998, "finished": true, "Ev Idx": 10828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564851968.777, "dur": 15.787,
    "args": {
      "Python parent id": 3990, "Python id": 3999, "finished": true, "Ev Idx": 10829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564851968.954, "dur": 0.270,
    "args": {
      "Python parent id": 3999, "Python id": 4000, "finished": true, "Ev Idx": 10830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564851969.864, "dur": 14.557,
    "args": {
      "Python parent id": 3999, "Python id": 4001, "finished": true, "Ev Idx": 10831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564851985.512, "dur": 1.762,
    "args": {
      "Python parent id": 3990, "Python id": 4002, "finished": true, "Ev Idx": 10832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564851988.710, "dur": 9.772,
    "args": {
      "Python parent id": 3990, "Python id": 4003, "finished": true, "Ev Idx": 10833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564851988.833, "dur": 0.106,
    "args": {
      "Python parent id": 4003, "Python id": 4004, "finished": true, "Ev Idx": 10834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564851994.213, "dur": 1.092,
    "args": {
      "Python parent id": 4003, "Python id": 4005, "finished": true, "Ev Idx": 10835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564851994.741, "dur": 0.484,
    "args": {
      "Python parent id": 4005, "Python id": 4006, "finished": true, "Ev Idx": 10836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564851995.706, "dur": 2.695,
    "args": {
      "Python parent id": 4003, "Python id": 4007, "finished": true, "Ev Idx": 10837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564851999.182, "dur": 38.612,
    "args": {
      "Python parent id": 3990, "Python id": 4008, "finished": true, "Ev Idx": 10838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564852038.448, "dur": 3.102,
    "args": {
      "Python parent id": 3990, "Python id": 4009, "finished": true, "Ev Idx": 10839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564852042.157, "dur": 13.858,
    "args": {
      "Python parent id": 3990, "Python id": 4010, "finished": true, "Ev Idx": 10840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564852059.573, "dur": 3.072,
    "args": {
      "Python parent id": 3941, "Python id": 4011, "finished": true, "Ev Idx": 10841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564852063.331, "dur": 0.320,
    "args": {
      "Python parent id": 3941, "Python id": 4012, "finished": true, "Ev Idx": 10842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852064.698, "dur": 0.845,
    "args": {
      "Python parent id": 3941, "Python id": 4013, "finished": true, "Ev Idx": 10843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_171", "pid": 537, "tid": 537,
    "ts": 7429564852069.525, "dur": 43.727,
    "args": {
      "Python parent id": 3941, "Python module id": 171, "Python id": 4014, "finished": true, "Ev Idx": 10844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852070.483, "dur": 42.388,
    "args": {
      "Python parent id": 4014, "Python id": 4015, "finished": true, "Ev Idx": 10845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852070.891, "dur": 1.210,
    "args": {
      "Python parent id": 4015, "Python id": 4016, "finished": true, "Ev Idx": 10846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564852073.213, "dur": 39.446,
    "args": {
      "Python parent id": 4015, "Python id": 4017, "finished": true, "Ev Idx": 10847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852073.648, "dur": 0.210,
    "args": {
      "Python parent id": 4017, "Python id": 4018, "finished": true, "Ev Idx": 10848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852074.105, "dur": 0.169,
    "args": {
      "Python parent id": 4017, "Python id": 4019, "finished": true, "Ev Idx": 10849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564852074.463, "dur": 38.098,
    "args": {
      "Python parent id": 4017, "Python id": 4020, "finished": true, "Ev Idx": 10850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852127.654, "dur": 0.739,
    "args": {
      "Python parent id": 3925, "Python id": 4021, "finished": true, "Ev Idx": 10851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_49", "pid": 537, "tid": 537,
    "ts": 7429564852131.102, "dur": 80.566,
    "args": {
      "Python parent id": 3925, "Python module id": 49, "Python id": 4022, "finished": true, "Ev Idx": 10852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852131.739, "dur": 79.458,
    "args": {
      "Python parent id": 4022, "Python id": 4023, "finished": true, "Ev Idx": 10853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852131.895, "dur": 0.447,
    "args": {
      "Python parent id": 4023, "Python id": 4024, "finished": true, "Ev Idx": 10854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564852133.358, "dur": 77.039,
    "args": {
      "Python parent id": 4023, "Python id": 4025, "finished": true, "Ev Idx": 10855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564852133.831, "dur": 1.629,
    "args": {
      "Python parent id": 4025, "Python id": 4026, "finished": true, "Ev Idx": 10856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564852135.742, "dur": 13.124,
    "args": {
      "Python parent id": 4025, "Python id": 4027, "finished": true, "Ev Idx": 10857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564852149.207, "dur": 13.594,
    "args": {
      "Python parent id": 4025, "Python id": 4028, "finished": true, "Ev Idx": 10858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564852176.456, "dur": 12.875,
    "args": {
      "Python parent id": 4025, "Python id": 4029, "finished": true, "Ev Idx": 10859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852199.348, "dur": 0.249,
    "args": {
      "Python parent id": 4025, "Python id": 4030, "finished": true, "Ev Idx": 10860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564852199.960, "dur": 1.274,
    "args": {
      "Python parent id": 4025, "Python id": 4031, "finished": true, "Ev Idx": 10861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852212.431, "dur": 0.335,
    "args": {
      "Python parent id": 3925, "Python id": 4032, "finished": true, "Ev Idx": 10862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_24", "pid": 537, "tid": 537,
    "ts": 7429564852214.893, "dur": 277.506,
    "args": {
      "Python parent id": 3925, "Python module id": 24, "Python id": 4033, "finished": true, "Ev Idx": 10863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852215.554, "dur": 276.609,
    "args": {
      "Python parent id": 4033, "Python id": 4034, "finished": true, "Ev Idx": 10864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852215.783, "dur": 0.414,
    "args": {
      "Python parent id": 4034, "Python id": 4035, "finished": true, "Ev Idx": 10865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564852217.044, "dur": 274.959,
    "args": {
      "Python parent id": 4034, "Python id": 4036, "finished": true, "Ev Idx": 10866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852217.430, "dur": 0.504,
    "args": {
      "Python parent id": 4036, "Python id": 4037, "finished": true, "Ev Idx": 10867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852218.285, "dur": 0.315,
    "args": {
      "Python parent id": 4036, "Python id": 4038, "finished": true, "Ev Idx": 10868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852218.873, "dur": 0.270,
    "args": {
      "Python parent id": 4036, "Python id": 4039, "finished": true, "Ev Idx": 10869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_172", "pid": 537, "tid": 537,
    "ts": 7429564852221.252, "dur": 52.599,
    "args": {
      "Python parent id": 4036, "Python module id": 172, "Python id": 4040, "finished": true, "Ev Idx": 10870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852221.908, "dur": 51.645,
    "args": {
      "Python parent id": 4040, "Python id": 4041, "finished": true, "Ev Idx": 10871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852222.160, "dur": 0.170,
    "args": {
      "Python parent id": 4041, "Python id": 4042, "finished": true, "Ev Idx": 10872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564852222.897, "dur": 50.473,
    "args": {
      "Python parent id": 4041, "Python id": 4043, "finished": true, "Ev Idx": 10873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852223.167, "dur": 0.139,
    "args": {
      "Python parent id": 4043, "Python id": 4044, "finished": true, "Ev Idx": 10874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852223.496, "dur": 0.285,
    "args": {
      "Python parent id": 4043, "Python id": 4045, "finished": true, "Ev Idx": 10875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564852223.939, "dur": 49.326,
    "args": {
      "Python parent id": 4043, "Python id": 4046, "finished": true, "Ev Idx": 10876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_24", "pid": 537, "tid": 537,
    "ts": 7429564852276.125, "dur": 68.862,
    "args": {
      "Python parent id": 4036, "Python module id": 24, "Python id": 4047, "finished": true, "Ev Idx": 10877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852276.618, "dur": 68.029,
    "args": {
      "Python parent id": 4047, "Python id": 4048, "finished": true, "Ev Idx": 10878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852276.841, "dur": 0.380,
    "args": {
      "Python parent id": 4048, "Python id": 4049, "finished": true, "Ev Idx": 10879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564852278.183, "dur": 66.371,
    "args": {
      "Python parent id": 4048, "Python id": 4050, "finished": true, "Ev Idx": 10880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564852279.091, "dur": 65.237,
    "args": {
      "Python parent id": 4050, "Python id": 4051, "finished": true, "Ev Idx": 10881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564852279.231, "dur": 0.218,
    "args": {
      "Python parent id": 4051, "Python id": 4052, "finished": true, "Ev Idx": 10882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564852280.412, "dur": 63.763,
    "args": {
      "Python parent id": 4051, "Python id": 4053, "finished": true, "Ev Idx": 10883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852345.957, "dur": 0.410,
    "args": {
      "Python parent id": 4036, "Python id": 4054, "finished": true, "Ev Idx": 10884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_173", "pid": 537, "tid": 537,
    "ts": 7429564852348.520, "dur": 49.653,
    "args": {
      "Python parent id": 4036, "Python module id": 173, "Python id": 4055, "finished": true, "Ev Idx": 10885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852349.144, "dur": 48.830,
    "args": {
      "Python parent id": 4055, "Python id": 4056, "finished": true, "Ev Idx": 10886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852349.307, "dur": 0.345,
    "args": {
      "Python parent id": 4056, "Python id": 4057, "finished": true, "Ev Idx": 10887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564852350.315, "dur": 47.538,
    "args": {
      "Python parent id": 4056, "Python id": 4058, "finished": true, "Ev Idx": 10888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852350.615, "dur": 0.150,
    "args": {
      "Python parent id": 4058, "Python id": 4059, "finished": true, "Ev Idx": 10889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852350.963, "dur": 0.120,
    "args": {
      "Python parent id": 4058, "Python id": 4060, "finished": true, "Ev Idx": 10890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564852351.190, "dur": 46.608,
    "args": {
      "Python parent id": 4058, "Python id": 4061, "finished": true, "Ev Idx": 10891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_174", "pid": 537, "tid": 537,
    "ts": 7429564852443.693, "dur": 48.013,
    "args": {
      "Python parent id": 4036, "Python module id": 174, "Python id": 4062, "finished": true, "Ev Idx": 10892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852444.548, "dur": 46.953,
    "args": {
      "Python parent id": 4062, "Python id": 4063, "finished": true, "Ev Idx": 10893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852444.764, "dur": 0.453,
    "args": {
      "Python parent id": 4063, "Python id": 4064, "finished": true, "Ev Idx": 10894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564852445.856, "dur": 45.503,
    "args": {
      "Python parent id": 4063, "Python id": 4065, "finished": true, "Ev Idx": 10895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852446.239, "dur": 0.170,
    "args": {
      "Python parent id": 4065, "Python id": 4066, "finished": true, "Ev Idx": 10896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852446.626, "dur": 0.139,
    "args": {
      "Python parent id": 4065, "Python id": 4067, "finished": true, "Ev Idx": 10897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564852446.886, "dur": 44.407,
    "args": {
      "Python parent id": 4065, "Python id": 4068, "finished": true, "Ev Idx": 10898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_25", "pid": 537, "tid": 537,
    "ts": 7429564852512.286, "dur": 1168.558,
    "args": {
      "Python parent id": 64, "Python module id": 25, "Python id": 4069, "finished": true, "Ev Idx": 10899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852513.793, "dur": 1166.208,
    "args": {
      "Python parent id": 4069, "Python id": 4070, "finished": true, "Ev Idx": 10900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852514.057, "dur": 0.433,
    "args": {
      "Python parent id": 4070, "Python id": 4071, "finished": true, "Ev Idx": 10901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564852516.194, "dur": 1162.732,
    "args": {
      "Python parent id": 4070, "Python id": 4072, "finished": true, "Ev Idx": 10902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852516.727, "dur": 0.471,
    "args": {
      "Python parent id": 4072, "Python id": 4073, "finished": true, "Ev Idx": 10903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_50", "pid": 537, "tid": 537,
    "ts": 7429564852519.797, "dur": 80.310,
    "args": {
      "Python parent id": 4072, "Python module id": 50, "Python id": 4074, "finished": true, "Ev Idx": 10904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852520.412, "dur": 79.270,
    "args": {
      "Python parent id": 4074, "Python id": 4075, "finished": true, "Ev Idx": 10905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852520.562, "dur": 0.157,
    "args": {
      "Python parent id": 4075, "Python id": 4076, "finished": true, "Ev Idx": 10906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564852521.454, "dur": 77.383,
    "args": {
      "Python parent id": 4075, "Python id": 4077, "finished": true, "Ev Idx": 10907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564852522.029, "dur": 2.110,
    "args": {
      "Python parent id": 4077, "Python id": 4078, "finished": true, "Ev Idx": 10908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564852524.454, "dur": 13.867,
    "args": {
      "Python parent id": 4077, "Python id": 4079, "finished": true, "Ev Idx": 10909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564852538.740, "dur": 13.904,
    "args": {
      "Python parent id": 4077, "Python id": 4080, "finished": true, "Ev Idx": 10910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564852565.335, "dur": 10.159,
    "args": {
      "Python parent id": 4077, "Python id": 4081, "finished": true, "Ev Idx": 10911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852584.875, "dur": 0.359,
    "args": {
      "Python parent id": 4077, "Python id": 4082, "finished": true, "Ev Idx": 10912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564852585.897, "dur": 1.348,
    "args": {
      "Python parent id": 4077, "Python id": 4083, "finished": true, "Ev Idx": 10913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852600.779, "dur": 0.410,
    "args": {
      "Python parent id": 4072, "Python id": 4084, "finished": true, "Ev Idx": 10914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_25", "pid": 537, "tid": 537,
    "ts": 7429564852604.000, "dur": 650.603,
    "args": {
      "Python parent id": 4072, "Python module id": 25, "Python id": 4085, "finished": true, "Ev Idx": 10915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852605.939, "dur": 648.160,
    "args": {
      "Python parent id": 4085, "Python id": 4086, "finished": true, "Ev Idx": 10916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852606.139, "dur": 0.336,
    "args": {
      "Python parent id": 4086, "Python id": 4087, "finished": true, "Ev Idx": 10917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564852608.289, "dur": 644.957,
    "args": {
      "Python parent id": 4086, "Python id": 4088, "finished": true, "Ev Idx": 10918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852611.567, "dur": 0.376,
    "args": {
      "Python parent id": 4088, "Python id": 4089, "finished": true, "Ev Idx": 10919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_175", "pid": 537, "tid": 537,
    "ts": 7429564852614.224, "dur": 43.919,
    "args": {
      "Python parent id": 4088, "Python module id": 175, "Python id": 4090, "finished": true, "Ev Idx": 10920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852614.729, "dur": 43.090,
    "args": {
      "Python parent id": 4090, "Python id": 4091, "finished": true, "Ev Idx": 10921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852614.883, "dur": 0.178,
    "args": {
      "Python parent id": 4091, "Python id": 4092, "finished": true, "Ev Idx": 10922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564852615.830, "dur": 41.848,
    "args": {
      "Python parent id": 4091, "Python id": 4093, "finished": true, "Ev Idx": 10923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852616.193, "dur": 0.165,
    "args": {
      "Python parent id": 4093, "Python id": 4094, "finished": true, "Ev Idx": 10924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852616.674, "dur": 0.149,
    "args": {
      "Python parent id": 4093, "Python id": 4095, "finished": true, "Ev Idx": 10925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564852617.005, "dur": 40.628,
    "args": {
      "Python parent id": 4093, "Python id": 4096, "finished": true, "Ev Idx": 10926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564852658.779, "dur": 3.504,
    "args": {
      "Python parent id": 4088, "Python id": 4097, "finished": true, "Ev Idx": 10927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564852662.756, "dur": 3.787,
    "args": {
      "Python parent id": 4088, "Python id": 4098, "finished": true, "Ev Idx": 10928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852667.343, "dur": 0.344,
    "args": {
      "Python parent id": 4088, "Python id": 4099, "finished": true, "Ev Idx": 10929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_176", "pid": 537, "tid": 537,
    "ts": 7429564852669.945, "dur": 43.783,
    "args": {
      "Python parent id": 4088, "Python module id": 176, "Python id": 4100, "finished": true, "Ev Idx": 10930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852670.530, "dur": 42.991,
    "args": {
      "Python parent id": 4100, "Python id": 4101, "finished": true, "Ev Idx": 10931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852670.767, "dur": 0.278,
    "args": {
      "Python parent id": 4101, "Python id": 4102, "finished": true, "Ev Idx": 10932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564852671.737, "dur": 41.656,
    "args": {
      "Python parent id": 4101, "Python id": 4103, "finished": true, "Ev Idx": 10933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852671.965, "dur": 0.151,
    "args": {
      "Python parent id": 4103, "Python id": 4104, "finished": true, "Ev Idx": 10934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852672.387, "dur": 0.149,
    "args": {
      "Python parent id": 4103, "Python id": 4105, "finished": true, "Ev Idx": 10935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564852672.651, "dur": 40.694,
    "args": {
      "Python parent id": 4103, "Python id": 4106, "finished": true, "Ev Idx": 10936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564852714.201, "dur": 2.375,
    "args": {
      "Python parent id": 4088, "Python id": 4107, "finished": true, "Ev Idx": 10937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564852717.125, "dur": 4.826,
    "args": {
      "Python parent id": 4088, "Python id": 4108, "finished": true, "Ev Idx": 10938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852722.761, "dur": 0.486,
    "args": {
      "Python parent id": 4088, "Python id": 4109, "finished": true, "Ev Idx": 10939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_177", "pid": 537, "tid": 537,
    "ts": 7429564852725.399, "dur": 41.912,
    "args": {
      "Python parent id": 4088, "Python module id": 177, "Python id": 4110, "finished": true, "Ev Idx": 10940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564852725.868, "dur": 41.236,
    "args": {
      "Python parent id": 4110, "Python id": 4111, "finished": true, "Ev Idx": 10941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564852726.072, "dur": 0.253,
    "args": {
      "Python parent id": 4111, "Python id": 4112, "finished": true, "Ev Idx": 10942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564852726.904, "dur": 40.090,
    "args": {
      "Python parent id": 4111, "Python id": 4113, "finished": true, "Ev Idx": 10943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852727.114, "dur": 0.127,
    "args": {
      "Python parent id": 4113, "Python id": 4114, "finished": true, "Ev Idx": 10944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564852727.447, "dur": 0.146,
    "args": {
      "Python parent id": 4113, "Python id": 4115, "finished": true, "Ev Idx": 10945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564852727.711, "dur": 39.235,
    "args": {
      "Python parent id": 4113, "Python id": 4116, "finished": true, "Ev Idx": 10946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564852767.826, "dur": 2.360,
    "args": {
      "Python parent id": 4088, "Python id": 4117, "finished": true, "Ev Idx": 10947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564852770.646, "dur": 2.657,
    "args": {
      "Python parent id": 4088, "Python id": 4118, "finished": true, "Ev Idx": 10948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564852774.414, "dur": 148.395,
    "args": {
      "Python parent id": 4088, "Python id": 4119, "finished": true, "Ev Idx": 10949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564852774.724, "dur": 3.650,
    "args": {
      "Python parent id": 4119, "Python id": 4120, "finished": true, "Ev Idx": 10950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564852778.712, "dur": 1.983,
    "args": {
      "Python parent id": 4119, "Python id": 4121, "finished": true, "Ev Idx": 10951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564852792.670, "dur": 49.849,
    "args": {
      "Python parent id": 4119, "Python id": 4122, "finished": true, "Ev Idx": 10952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564852820.718, "dur": 20.619,
    "args": {
      "Python parent id": 4122, "Python id": 4123, "finished": true, "Ev Idx": 10953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564852872.644, "dur": 30.717,
    "args": {
      "Python parent id": 4119, "Python id": 4124, "finished": true, "Ev Idx": 10954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564852889.389, "dur": 12.893,
    "args": {
      "Python parent id": 4124, "Python id": 4125, "finished": true, "Ev Idx": 10955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564852927.032, "dur": 3.775,
    "args": {
      "Python parent id": 4088, "Python id": 4126, "finished": true, "Ev Idx": 10956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564852927.596, "dur": 0.270,
    "args": {
      "Python parent id": 4126, "Python id": 4127, "finished": true, "Ev Idx": 10957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564852928.320, "dur": 0.056,
    "args": {
      "Python parent id": 4126, "Python id": 4128, "finished": true, "Ev Idx": 10958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564852929.799, "dur": 0.232,
    "args": {
      "Python parent id": 4126, "Python id": 4129, "finished": true, "Ev Idx": 10959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564852930.434, "dur": 0.135,
    "args": {
      "Python parent id": 4126, "Python id": 4130, "finished": true, "Ev Idx": 10960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564852932.174, "dur": 6.537,
    "args": {
      "Python parent id": 4088, "Python id": 4131, "finished": true, "Ev Idx": 10961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564852935.099, "dur": 3.478,
    "args": {
      "Python parent id": 4131, "Python id": 4132, "finished": true, "Ev Idx": 10962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564852935.350, "dur": 1.553,
    "args": {
      "Python parent id": 4132, "Python id": 4133, "finished": true, "Ev Idx": 10963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564852935.954, "dur": 0.786,
    "args": {
      "Python parent id": 4133, "Python id": 4134, "finished": true, "Ev Idx": 10964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564852937.220, "dur": 0.562,
    "args": {
      "Python parent id": 4132, "Python id": 4135, "finished": true, "Ev Idx": 10965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564852938.080, "dur": 0.422,
    "args": {
      "Python parent id": 4132, "Python id": 4136, "finished": true, "Ev Idx": 10966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564852941.458, "dur": 254.457,
    "args": {
      "Python parent id": 4088, "Python id": 4137, "finished": true, "Ev Idx": 10967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564852942.041, "dur": 46.196,
    "args": {
      "Python parent id": 4137, "Python id": 4138, "finished": true, "Ev Idx": 10968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564852960.201, "dur": 5.769,
    "args": {
      "Python parent id": 4138, "Python id": 4139, "finished": true, "Ev Idx": 10969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564852966.819, "dur": 21.272,
    "args": {
      "Python parent id": 4138, "Python id": 4140, "finished": true, "Ev Idx": 10970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564852989.143, "dur": 32.540,
    "args": {
      "Python parent id": 4137, "Python id": 4141, "finished": true, "Ev Idx": 10971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564853004.656, "dur": 2.823,
    "args": {
      "Python parent id": 4141, "Python id": 4142, "finished": true, "Ev Idx": 10972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564853007.947, "dur": 13.632,
    "args": {
      "Python parent id": 4141, "Python id": 4143, "finished": true, "Ev Idx": 10973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564853022.752, "dur": 3.955,
    "args": {
      "Python parent id": 4137, "Python id": 4144, "finished": true, "Ev Idx": 10974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564853026.992, "dur": 45.686,
    "args": {
      "Python parent id": 4137, "Python id": 4145, "finished": true, "Ev Idx": 10975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564853104.791, "dur": 16.751,
    "args": {
      "Python parent id": 4137, "Python id": 4146, "finished": true, "Ev Idx": 10976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564853104.948, "dur": 0.264,
    "args": {
      "Python parent id": 4146, "Python id": 4147, "finished": true, "Ev Idx": 10977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564853105.899, "dur": 15.465,
    "args": {
      "Python parent id": 4146, "Python id": 4148, "finished": true, "Ev Idx": 10978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564853122.487, "dur": 1.760,
    "args": {
      "Python parent id": 4137, "Python id": 4149, "finished": true, "Ev Idx": 10979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564853125.669, "dur": 10.146,
    "args": {
      "Python parent id": 4137, "Python id": 4150, "finished": true, "Ev Idx": 10980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564853125.775, "dur": 0.103,
    "args": {
      "Python parent id": 4150, "Python id": 4151, "finished": true, "Ev Idx": 10981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853131.401, "dur": 0.997,
    "args": {
      "Python parent id": 4150, "Python id": 4152, "finished": true, "Ev Idx": 10982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564853131.887, "dur": 0.424,
    "args": {
      "Python parent id": 4152, "Python id": 4153, "finished": true, "Ev Idx": 10983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564853132.809, "dur": 2.915,
    "args": {
      "Python parent id": 4150, "Python id": 4154, "finished": true, "Ev Idx": 10984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564853136.474, "dur": 40.091,
    "args": {
      "Python parent id": 4137, "Python id": 4155, "finished": true, "Ev Idx": 10985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564853177.260, "dur": 2.976,
    "args": {
      "Python parent id": 4137, "Python id": 4156, "finished": true, "Ev Idx": 10986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564853180.887, "dur": 14.087,
    "args": {
      "Python parent id": 4137, "Python id": 4157, "finished": true, "Ev Idx": 10987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564853198.542, "dur": 3.097,
    "args": {
      "Python parent id": 4088, "Python id": 4158, "finished": true, "Ev Idx": 10988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564853202.283, "dur": 0.261,
    "args": {
      "Python parent id": 4088, "Python id": 4159, "finished": true, "Ev Idx": 10989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853203.505, "dur": 0.796,
    "args": {
      "Python parent id": 4088, "Python id": 4160, "finished": true, "Ev Idx": 10990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_178", "pid": 537, "tid": 537,
    "ts": 7429564853207.766, "dur": 44.885,
    "args": {
      "Python parent id": 4088, "Python module id": 178, "Python id": 4161, "finished": true, "Ev Idx": 10991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853208.587, "dur": 43.731,
    "args": {
      "Python parent id": 4161, "Python id": 4162, "finished": true, "Ev Idx": 10992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853208.983, "dur": 0.771,
    "args": {
      "Python parent id": 4162, "Python id": 4163, "finished": true, "Ev Idx": 10993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564853210.823, "dur": 41.246,
    "args": {
      "Python parent id": 4162, "Python id": 4164, "finished": true, "Ev Idx": 10994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853211.213, "dur": 0.188,
    "args": {
      "Python parent id": 4164, "Python id": 4165, "finished": true, "Ev Idx": 10995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853211.661, "dur": 0.136,
    "args": {
      "Python parent id": 4164, "Python id": 4166, "finished": true, "Ev Idx": 10996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564853211.934, "dur": 40.036,
    "args": {
      "Python parent id": 4164, "Python id": 4167, "finished": true, "Ev Idx": 10997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853324.696, "dur": 0.722,
    "args": {
      "Python parent id": 4072, "Python id": 4168, "finished": true, "Ev Idx": 10998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_51", "pid": 537, "tid": 537,
    "ts": 7429564853328.259, "dur": 89.379,
    "args": {
      "Python parent id": 4072, "Python module id": 51, "Python id": 4169, "finished": true, "Ev Idx": 10999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853329.041, "dur": 88.228,
    "args": {
      "Python parent id": 4169, "Python id": 4170, "finished": true, "Ev Idx": 11000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853329.376, "dur": 0.670,
    "args": {
      "Python parent id": 4170, "Python id": 4171, "finished": true, "Ev Idx": 11001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564853331.484, "dur": 84.822,
    "args": {
      "Python parent id": 4170, "Python id": 4172, "finished": true, "Ev Idx": 11002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564853332.074, "dur": 2.211,
    "args": {
      "Python parent id": 4172, "Python id": 4173, "finished": true, "Ev Idx": 11003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564853334.616, "dur": 14.786,
    "args": {
      "Python parent id": 4172, "Python id": 4174, "finished": true, "Ev Idx": 11004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564853349.780, "dur": 16.962,
    "args": {
      "Python parent id": 4172, "Python id": 4175, "finished": true, "Ev Idx": 11005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564853379.386, "dur": 11.542,
    "args": {
      "Python parent id": 4172, "Python id": 4176, "finished": true, "Ev Idx": 11006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853403.790, "dur": 0.203,
    "args": {
      "Python parent id": 4172, "Python id": 4177, "finished": true, "Ev Idx": 11007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564853404.335, "dur": 1.331,
    "args": {
      "Python parent id": 4172, "Python id": 4178, "finished": true, "Ev Idx": 11008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853418.395, "dur": 0.328,
    "args": {
      "Python parent id": 4072, "Python id": 4179, "finished": true, "Ev Idx": 11009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_25", "pid": 537, "tid": 537,
    "ts": 7429564853420.431, "dur": 247.401,
    "args": {
      "Python parent id": 4072, "Python module id": 25, "Python id": 4180, "finished": true, "Ev Idx": 11010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853420.994, "dur": 246.611,
    "args": {
      "Python parent id": 4180, "Python id": 4181, "finished": true, "Ev Idx": 11011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853421.262, "dur": 0.491,
    "args": {
      "Python parent id": 4181, "Python id": 4182, "finished": true, "Ev Idx": 11012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564853422.701, "dur": 244.791,
    "args": {
      "Python parent id": 4181, "Python id": 4183, "finished": true, "Ev Idx": 11013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853423.037, "dur": 0.405,
    "args": {
      "Python parent id": 4183, "Python id": 4184, "finished": true, "Ev Idx": 11014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853423.739, "dur": 0.301,
    "args": {
      "Python parent id": 4183, "Python id": 4185, "finished": true, "Ev Idx": 11015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853424.291, "dur": 0.202,
    "args": {
      "Python parent id": 4183, "Python id": 4186, "finished": true, "Ev Idx": 11016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_179", "pid": 537, "tid": 537,
    "ts": 7429564853426.694, "dur": 57.253,
    "args": {
      "Python parent id": 4183, "Python module id": 179, "Python id": 4187, "finished": true, "Ev Idx": 11017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853427.233, "dur": 56.424,
    "args": {
      "Python parent id": 4187, "Python id": 4188, "finished": true, "Ev Idx": 11018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853427.370, "dur": 0.199,
    "args": {
      "Python parent id": 4188, "Python id": 4189, "finished": true, "Ev Idx": 11019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564853428.134, "dur": 55.398,
    "args": {
      "Python parent id": 4188, "Python id": 4190, "finished": true, "Ev Idx": 11020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853428.468, "dur": 0.118,
    "args": {
      "Python parent id": 4190, "Python id": 4191, "finished": true, "Ev Idx": 11021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853428.806, "dur": 0.114,
    "args": {
      "Python parent id": 4190, "Python id": 4192, "finished": true, "Ev Idx": 11022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564853429.230, "dur": 54.198,
    "args": {
      "Python parent id": 4190, "Python id": 4193, "finished": true, "Ev Idx": 11023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_25", "pid": 537, "tid": 537,
    "ts": 7429564853485.699, "dur": 16.074,
    "args": {
      "Python parent id": 4183, "Python module id": 25, "Python id": 4194, "finished": true, "Ev Idx": 11024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853486.153, "dur": 15.419,
    "args": {
      "Python parent id": 4194, "Python id": 4195, "finished": true, "Ev Idx": 11025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853486.333, "dur": 0.327,
    "args": {
      "Python parent id": 4195, "Python id": 4196, "finished": true, "Ev Idx": 11026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564853487.620, "dur": 13.869,
    "args": {
      "Python parent id": 4195, "Python id": 4197, "finished": true, "Ev Idx": 11027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564853488.574, "dur": 12.777,
    "args": {
      "Python parent id": 4197, "Python id": 4198, "finished": true, "Ev Idx": 11028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564853488.729, "dur": 0.243,
    "args": {
      "Python parent id": 4198, "Python id": 4199, "finished": true, "Ev Idx": 11029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564853489.886, "dur": 11.375,
    "args": {
      "Python parent id": 4198, "Python id": 4200, "finished": true, "Ev Idx": 11030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853502.504, "dur": 0.351,
    "args": {
      "Python parent id": 4183, "Python id": 4201, "finished": true, "Ev Idx": 11031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_180", "pid": 537, "tid": 537,
    "ts": 7429564853504.850, "dur": 91.301,
    "args": {
      "Python parent id": 4183, "Python module id": 180, "Python id": 4202, "finished": true, "Ev Idx": 11032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853505.463, "dur": 90.501,
    "args": {
      "Python parent id": 4202, "Python id": 4203, "finished": true, "Ev Idx": 11033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853505.676, "dur": 0.203,
    "args": {
      "Python parent id": 4203, "Python id": 4204, "finished": true, "Ev Idx": 11034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564853506.480, "dur": 89.305,
    "args": {
      "Python parent id": 4203, "Python id": 4205, "finished": true, "Ev Idx": 11035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853506.767, "dur": 0.119,
    "args": {
      "Python parent id": 4205, "Python id": 4206, "finished": true, "Ev Idx": 11036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853507.089, "dur": 0.145,
    "args": {
      "Python parent id": 4205, "Python id": 4207, "finished": true, "Ev Idx": 11037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564853507.335, "dur": 88.323,
    "args": {
      "Python parent id": 4205, "Python id": 4208, "finished": true, "Ev Idx": 11038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_181", "pid": 537, "tid": 537,
    "ts": 7429564853609.877, "dur": 57.362,
    "args": {
      "Python parent id": 4183, "Python module id": 181, "Python id": 4209, "finished": true, "Ev Idx": 11039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853610.518, "dur": 56.478,
    "args": {
      "Python parent id": 4209, "Python id": 4210, "finished": true, "Ev Idx": 11040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853610.806, "dur": 0.424,
    "args": {
      "Python parent id": 4210, "Python id": 4211, "finished": true, "Ev Idx": 11041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564853611.939, "dur": 54.924,
    "args": {
      "Python parent id": 4210, "Python id": 4212, "finished": true, "Ev Idx": 11042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853612.265, "dur": 0.235,
    "args": {
      "Python parent id": 4212, "Python id": 4213, "finished": true, "Ev Idx": 11043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853612.721, "dur": 0.115,
    "args": {
      "Python parent id": 4212, "Python id": 4214, "finished": true, "Ev Idx": 11044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564853612.943, "dur": 53.849,
    "args": {
      "Python parent id": 4212, "Python id": 4215, "finished": true, "Ev Idx": 11045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_26", "pid": 537, "tid": 537,
    "ts": 7429564853686.927, "dur": 1247.798,
    "args": {
      "Python parent id": 64, "Python module id": 26, "Python id": 4216, "finished": true, "Ev Idx": 11046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853688.248, "dur": 1245.558,
    "args": {
      "Python parent id": 4216, "Python id": 4217, "finished": true, "Ev Idx": 11047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853688.498, "dur": 0.278,
    "args": {
      "Python parent id": 4217, "Python id": 4218, "finished": true, "Ev Idx": 11048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564853690.416, "dur": 1242.464,
    "args": {
      "Python parent id": 4217, "Python id": 4219, "finished": true, "Ev Idx": 11049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853690.901, "dur": 0.574,
    "args": {
      "Python parent id": 4219, "Python id": 4220, "finished": true, "Ev Idx": 11050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_52", "pid": 537, "tid": 537,
    "ts": 7429564853693.637, "dur": 106.354,
    "args": {
      "Python parent id": 4219, "Python module id": 52, "Python id": 4221, "finished": true, "Ev Idx": 11051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853694.229, "dur": 105.345,
    "args": {
      "Python parent id": 4221, "Python id": 4222, "finished": true, "Ev Idx": 11052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853694.500, "dur": 0.166,
    "args": {
      "Python parent id": 4222, "Python id": 4223, "finished": true, "Ev Idx": 11053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564853695.399, "dur": 103.339,
    "args": {
      "Python parent id": 4222, "Python id": 4224, "finished": true, "Ev Idx": 11054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564853695.983, "dur": 1.988,
    "args": {
      "Python parent id": 4224, "Python id": 4225, "finished": true, "Ev Idx": 11055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564853698.236, "dur": 11.053,
    "args": {
      "Python parent id": 4224, "Python id": 4226, "finished": true, "Ev Idx": 11056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564853709.609, "dur": 12.380,
    "args": {
      "Python parent id": 4224, "Python id": 4227, "finished": true, "Ev Idx": 11057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564853734.412, "dur": 9.503,
    "args": {
      "Python parent id": 4224, "Python id": 4228, "finished": true, "Ev Idx": 11058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853753.084, "dur": 0.234,
    "args": {
      "Python parent id": 4224, "Python id": 4229, "finished": true, "Ev Idx": 11059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564853753.685, "dur": 1.597,
    "args": {
      "Python parent id": 4224, "Python id": 4230, "finished": true, "Ev Idx": 11060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853800.785, "dur": 0.403,
    "args": {
      "Python parent id": 4219, "Python id": 4231, "finished": true, "Ev Idx": 11061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_26", "pid": 537, "tid": 537,
    "ts": 7429564853803.902, "dur": 717.182,
    "args": {
      "Python parent id": 4219, "Python module id": 26, "Python id": 4232, "finished": true, "Ev Idx": 11062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853805.276, "dur": 715.245,
    "args": {
      "Python parent id": 4232, "Python id": 4233, "finished": true, "Ev Idx": 11063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853805.488, "dur": 0.391,
    "args": {
      "Python parent id": 4233, "Python id": 4234, "finished": true, "Ev Idx": 11064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564853807.482, "dur": 712.190,
    "args": {
      "Python parent id": 4233, "Python id": 4235, "finished": true, "Ev Idx": 11065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853811.059, "dur": 0.332,
    "args": {
      "Python parent id": 4235, "Python id": 4236, "finished": true, "Ev Idx": 11066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_182", "pid": 537, "tid": 537,
    "ts": 7429564853813.253, "dur": 41.143,
    "args": {
      "Python parent id": 4235, "Python module id": 182, "Python id": 4237, "finished": true, "Ev Idx": 11067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853813.798, "dur": 40.339,
    "args": {
      "Python parent id": 4237, "Python id": 4238, "finished": true, "Ev Idx": 11068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853813.979, "dur": 0.173,
    "args": {
      "Python parent id": 4238, "Python id": 4239, "finished": true, "Ev Idx": 11069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564853814.904, "dur": 39.069,
    "args": {
      "Python parent id": 4238, "Python id": 4240, "finished": true, "Ev Idx": 11070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853815.210, "dur": 0.165,
    "args": {
      "Python parent id": 4240, "Python id": 4241, "finished": true, "Ev Idx": 11071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853815.571, "dur": 0.128,
    "args": {
      "Python parent id": 4240, "Python id": 4242, "finished": true, "Ev Idx": 11072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564853815.798, "dur": 38.106,
    "args": {
      "Python parent id": 4240, "Python id": 4243, "finished": true, "Ev Idx": 11073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564853854.971, "dur": 3.496,
    "args": {
      "Python parent id": 4235, "Python id": 4244, "finished": true, "Ev Idx": 11074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564853858.941, "dur": 3.750,
    "args": {
      "Python parent id": 4235, "Python id": 4245, "finished": true, "Ev Idx": 11075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853863.435, "dur": 0.341,
    "args": {
      "Python parent id": 4235, "Python id": 4246, "finished": true, "Ev Idx": 11076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_183", "pid": 537, "tid": 537,
    "ts": 7429564853865.958, "dur": 43.006,
    "args": {
      "Python parent id": 4235, "Python module id": 183, "Python id": 4247, "finished": true, "Ev Idx": 11077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853866.504, "dur": 42.260,
    "args": {
      "Python parent id": 4247, "Python id": 4248, "finished": true, "Ev Idx": 11078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853866.738, "dur": 0.292,
    "args": {
      "Python parent id": 4248, "Python id": 4249, "finished": true, "Ev Idx": 11079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564853867.735, "dur": 40.938,
    "args": {
      "Python parent id": 4248, "Python id": 4250, "finished": true, "Ev Idx": 11080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853867.975, "dur": 0.165,
    "args": {
      "Python parent id": 4250, "Python id": 4251, "finished": true, "Ev Idx": 11081
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853868.353, "dur": 0.186,
    "args": {
      "Python parent id": 4250, "Python id": 4252, "finished": true, "Ev Idx": 11082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564853868.656, "dur": 39.959,
    "args": {
      "Python parent id": 4250, "Python id": 4253, "finished": true, "Ev Idx": 11083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564853909.430, "dur": 2.436,
    "args": {
      "Python parent id": 4235, "Python id": 4254, "finished": true, "Ev Idx": 11084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564853912.342, "dur": 2.854,
    "args": {
      "Python parent id": 4235, "Python id": 4255, "finished": true, "Ev Idx": 11085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853915.921, "dur": 0.448,
    "args": {
      "Python parent id": 4235, "Python id": 4256, "finished": true, "Ev Idx": 11086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_184", "pid": 537, "tid": 537,
    "ts": 7429564853918.737, "dur": 37.913,
    "args": {
      "Python parent id": 4235, "Python module id": 184, "Python id": 4257, "finished": true, "Ev Idx": 11087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564853919.382, "dur": 37.096,
    "args": {
      "Python parent id": 4257, "Python id": 4258, "finished": true, "Ev Idx": 11088
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564853919.658, "dur": 0.252,
    "args": {
      "Python parent id": 4258, "Python id": 4259, "finished": true, "Ev Idx": 11089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564853920.618, "dur": 35.747,
    "args": {
      "Python parent id": 4258, "Python id": 4260, "finished": true, "Ev Idx": 11090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853920.830, "dur": 0.152,
    "args": {
      "Python parent id": 4260, "Python id": 4261, "finished": true, "Ev Idx": 11091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564853921.154, "dur": 0.135,
    "args": {
      "Python parent id": 4260, "Python id": 4262, "finished": true, "Ev Idx": 11092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564853921.415, "dur": 34.902,
    "args": {
      "Python parent id": 4260, "Python id": 4263, "finished": true, "Ev Idx": 11093
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564853957.129, "dur": 2.510,
    "args": {
      "Python parent id": 4235, "Python id": 4264, "finished": true, "Ev Idx": 11094
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564853960.158, "dur": 2.464,
    "args": {
      "Python parent id": 4235, "Python id": 4265, "finished": true, "Ev Idx": 11095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564853963.708, "dur": 177.808,
    "args": {
      "Python parent id": 4235, "Python id": 4266, "finished": true, "Ev Idx": 11096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564853964.022, "dur": 3.843,
    "args": {
      "Python parent id": 4266, "Python id": 4267, "finished": true, "Ev Idx": 11097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564853968.146, "dur": 1.815,
    "args": {
      "Python parent id": 4266, "Python id": 4268, "finished": true, "Ev Idx": 11098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564853981.068, "dur": 46.337,
    "args": {
      "Python parent id": 4266, "Python id": 4269, "finished": true, "Ev Idx": 11099
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564854006.030, "dur": 20.204,
    "args": {
      "Python parent id": 4269, "Python id": 4270, "finished": true, "Ev Idx": 11100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564854058.746, "dur": 59.093,
    "args": {
      "Python parent id": 4266, "Python id": 4271, "finished": true, "Ev Idx": 11101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564854095.333, "dur": 21.422,
    "args": {
      "Python parent id": 4271, "Python id": 4272, "finished": true, "Ev Idx": 11102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564854146.244, "dur": 4.164,
    "args": {
      "Python parent id": 4235, "Python id": 4273, "finished": true, "Ev Idx": 11103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564854147.044, "dur": 0.393,
    "args": {
      "Python parent id": 4273, "Python id": 4274, "finished": true, "Ev Idx": 11104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564854147.865, "dur": 0.067,
    "args": {
      "Python parent id": 4273, "Python id": 4275, "finished": true, "Ev Idx": 11105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564854149.429, "dur": 0.271,
    "args": {
      "Python parent id": 4273, "Python id": 4276, "finished": true, "Ev Idx": 11106
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564854150.064, "dur": 0.105,
    "args": {
      "Python parent id": 4273, "Python id": 4277, "finished": true, "Ev Idx": 11107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564854151.819, "dur": 6.798,
    "args": {
      "Python parent id": 4235, "Python id": 4278, "finished": true, "Ev Idx": 11108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564854155.090, "dur": 3.389,
    "args": {
      "Python parent id": 4278, "Python id": 4279, "finished": true, "Ev Idx": 11109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564854155.300, "dur": 1.508,
    "args": {
      "Python parent id": 4279, "Python id": 4280, "finished": true, "Ev Idx": 11110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564854155.856, "dur": 0.768,
    "args": {
      "Python parent id": 4280, "Python id": 4281, "finished": true, "Ev Idx": 11111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564854157.087, "dur": 0.544,
    "args": {
      "Python parent id": 4279, "Python id": 4282, "finished": true, "Ev Idx": 11112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564854157.959, "dur": 0.444,
    "args": {
      "Python parent id": 4279, "Python id": 4283, "finished": true, "Ev Idx": 11113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564854161.381, "dur": 297.926,
    "args": {
      "Python parent id": 4235, "Python id": 4284, "finished": true, "Ev Idx": 11114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564854162.006, "dur": 75.905,
    "args": {
      "Python parent id": 4284, "Python id": 4285, "finished": true, "Ev Idx": 11115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564854180.889, "dur": 31.872,
    "args": {
      "Python parent id": 4285, "Python id": 4286, "finished": true, "Ev Idx": 11116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564854213.661, "dur": 24.087,
    "args": {
      "Python parent id": 4285, "Python id": 4287, "finished": true, "Ev Idx": 11117
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564854238.823, "dur": 30.247,
    "args": {
      "Python parent id": 4284, "Python id": 4288, "finished": true, "Ev Idx": 11118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564854252.914, "dur": 2.668,
    "args": {
      "Python parent id": 4288, "Python id": 4289, "finished": true, "Ev Idx": 11119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564854256.045, "dur": 12.960,
    "args": {
      "Python parent id": 4288, "Python id": 4290, "finished": true, "Ev Idx": 11120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564854270.054, "dur": 3.924,
    "args": {
      "Python parent id": 4284, "Python id": 4291, "finished": true, "Ev Idx": 11121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564854274.262, "dur": 54.092,
    "args": {
      "Python parent id": 4284, "Python id": 4292, "finished": true, "Ev Idx": 11122
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564854363.768, "dur": 18.910,
    "args": {
      "Python parent id": 4284, "Python id": 4293, "finished": true, "Ev Idx": 11123
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564854363.984, "dur": 0.340,
    "args": {
      "Python parent id": 4293, "Python id": 4294, "finished": true, "Ev Idx": 11124
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564854365.041, "dur": 17.443,
    "args": {
      "Python parent id": 4293, "Python id": 4295, "finished": true, "Ev Idx": 11125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564854383.652, "dur": 2.032,
    "args": {
      "Python parent id": 4284, "Python id": 4296, "finished": true, "Ev Idx": 11126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564854387.286, "dur": 10.101,
    "args": {
      "Python parent id": 4284, "Python id": 4297, "finished": true, "Ev Idx": 11127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564854387.442, "dur": 0.092,
    "args": {
      "Python parent id": 4297, "Python id": 4298, "finished": true, "Ev Idx": 11128
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854393.061, "dur": 0.994,
    "args": {
      "Python parent id": 4297, "Python id": 4299, "finished": true, "Ev Idx": 11129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564854393.522, "dur": 0.447,
    "args": {
      "Python parent id": 4299, "Python id": 4300, "finished": true, "Ev Idx": 11130
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564854394.488, "dur": 2.810,
    "args": {
      "Python parent id": 4297, "Python id": 4301, "finished": true, "Ev Idx": 11131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564854398.157, "dur": 41.714,
    "args": {
      "Python parent id": 4284, "Python id": 4302, "finished": true, "Ev Idx": 11132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564854440.548, "dur": 2.968,
    "args": {
      "Python parent id": 4284, "Python id": 4303, "finished": true, "Ev Idx": 11133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564854444.284, "dur": 14.004,
    "args": {
      "Python parent id": 4284, "Python id": 4304, "finished": true, "Ev Idx": 11134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564854462.129, "dur": 3.372,
    "args": {
      "Python parent id": 4235, "Python id": 4305, "finished": true, "Ev Idx": 11135
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564854466.158, "dur": 0.260,
    "args": {
      "Python parent id": 4235, "Python id": 4306, "finished": true, "Ev Idx": 11136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854467.540, "dur": 0.959,
    "args": {
      "Python parent id": 4235, "Python id": 4307, "finished": true, "Ev Idx": 11137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_185", "pid": 537, "tid": 537,
    "ts": 7429564854474.299, "dur": 44.669,
    "args": {
      "Python parent id": 4235, "Python module id": 185, "Python id": 4308, "finished": true, "Ev Idx": 11138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564854475.646, "dur": 42.832,
    "args": {
      "Python parent id": 4308, "Python id": 4309, "finished": true, "Ev Idx": 11139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564854476.140, "dur": 1.501,
    "args": {
      "Python parent id": 4309, "Python id": 4310, "finished": true, "Ev Idx": 11140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564854479.031, "dur": 39.198,
    "args": {
      "Python parent id": 4309, "Python id": 4311, "finished": true, "Ev Idx": 11141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854479.482, "dur": 0.357,
    "args": {
      "Python parent id": 4311, "Python id": 4312, "finished": true, "Ev Idx": 11142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854480.239, "dur": 0.192,
    "args": {
      "Python parent id": 4311, "Python id": 4313, "finished": true, "Ev Idx": 11143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564854480.700, "dur": 37.388,
    "args": {
      "Python parent id": 4311, "Python id": 4314, "finished": true, "Ev Idx": 11144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854533.380, "dur": 0.681,
    "args": {
      "Python parent id": 4219, "Python id": 4315, "finished": true, "Ev Idx": 11145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_53", "pid": 537, "tid": 537,
    "ts": 7429564854537.342, "dur": 84.808,
    "args": {
      "Python parent id": 4219, "Python module id": 53, "Python id": 4316, "finished": true, "Ev Idx": 11146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564854538.124, "dur": 83.558,
    "args": {
      "Python parent id": 4316, "Python id": 4317, "finished": true, "Ev Idx": 11147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564854538.419, "dur": 0.409,
    "args": {
      "Python parent id": 4317, "Python id": 4318, "finished": true, "Ev Idx": 11148
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564854539.640, "dur": 81.175,
    "args": {
      "Python parent id": 4317, "Python id": 4319, "finished": true, "Ev Idx": 11149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564854540.258, "dur": 1.790,
    "args": {
      "Python parent id": 4319, "Python id": 4320, "finished": true, "Ev Idx": 11150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564854542.391, "dur": 14.241,
    "args": {
      "Python parent id": 4319, "Python id": 4321, "finished": true, "Ev Idx": 11151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564854557.006, "dur": 16.482,
    "args": {
      "Python parent id": 4319, "Python id": 4322, "finished": true, "Ev Idx": 11152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564854586.920, "dur": 11.522,
    "args": {
      "Python parent id": 4319, "Python id": 4323, "finished": true, "Ev Idx": 11153
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854608.925, "dur": 0.339,
    "args": {
      "Python parent id": 4319, "Python id": 4324, "finished": true, "Ev Idx": 11154
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564854609.604, "dur": 1.261,
    "args": {
      "Python parent id": 4319, "Python id": 4325, "finished": true, "Ev Idx": 11155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854623.027, "dur": 0.386,
    "args": {
      "Python parent id": 4219, "Python id": 4326, "finished": true, "Ev Idx": 11156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_26", "pid": 537, "tid": 537,
    "ts": 7429564854625.460, "dur": 295.580,
    "args": {
      "Python parent id": 4219, "Python module id": 26, "Python id": 4327, "finished": true, "Ev Idx": 11157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564854626.076, "dur": 294.619,
    "args": {
      "Python parent id": 4327, "Python id": 4328, "finished": true, "Ev Idx": 11158
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564854626.290, "dur": 0.380,
    "args": {
      "Python parent id": 4328, "Python id": 4329, "finished": true, "Ev Idx": 11159
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564854627.564, "dur": 292.978,
    "args": {
      "Python parent id": 4328, "Python id": 4330, "finished": true, "Ev Idx": 11160
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854627.969, "dur": 0.437,
    "args": {
      "Python parent id": 4330, "Python id": 4331, "finished": true, "Ev Idx": 11161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854628.718, "dur": 0.351,
    "args": {
      "Python parent id": 4330, "Python id": 4332, "finished": true, "Ev Idx": 11162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854629.352, "dur": 0.250,
    "args": {
      "Python parent id": 4330, "Python id": 4333, "finished": true, "Ev Idx": 11163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_186", "pid": 537, "tid": 537,
    "ts": 7429564854631.619, "dur": 123.458,
    "args": {
      "Python parent id": 4330, "Python module id": 186, "Python id": 4334, "finished": true, "Ev Idx": 11164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564854632.147, "dur": 122.507,
    "args": {
      "Python parent id": 4334, "Python id": 4335, "finished": true, "Ev Idx": 11165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564854632.304, "dur": 0.162,
    "args": {
      "Python parent id": 4335, "Python id": 4336, "finished": true, "Ev Idx": 11166
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564854633.114, "dur": 121.303,
    "args": {
      "Python parent id": 4335, "Python id": 4337, "finished": true, "Ev Idx": 11167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854633.403, "dur": 0.129,
    "args": {
      "Python parent id": 4337, "Python id": 4338, "finished": true, "Ev Idx": 11168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854633.759, "dur": 0.146,
    "args": {
      "Python parent id": 4337, "Python id": 4339, "finished": true, "Ev Idx": 11169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564854634.078, "dur": 120.135,
    "args": {
      "Python parent id": 4337, "Python id": 4340, "finished": true, "Ev Idx": 11170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_26", "pid": 537, "tid": 537,
    "ts": 7429564854757.446, "dur": 19.081,
    "args": {
      "Python parent id": 4330, "Python module id": 26, "Python id": 4341, "finished": true, "Ev Idx": 11171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564854758.131, "dur": 18.149,
    "args": {
      "Python parent id": 4341, "Python id": 4342, "finished": true, "Ev Idx": 11172
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564854758.403, "dur": 0.466,
    "args": {
      "Python parent id": 4342, "Python id": 4343, "finished": true, "Ev Idx": 11173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564854760.076, "dur": 16.093,
    "args": {
      "Python parent id": 4342, "Python id": 4344, "finished": true, "Ev Idx": 11174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564854761.072, "dur": 14.916,
    "args": {
      "Python parent id": 4344, "Python id": 4345, "finished": true, "Ev Idx": 11175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564854761.220, "dur": 0.255,
    "args": {
      "Python parent id": 4345, "Python id": 4346, "finished": true, "Ev Idx": 11176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564854762.472, "dur": 13.421,
    "args": {
      "Python parent id": 4345, "Python id": 4347, "finished": true, "Ev Idx": 11177
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854777.364, "dur": 0.396,
    "args": {
      "Python parent id": 4330, "Python id": 4348, "finished": true, "Ev Idx": 11178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_187", "pid": 537, "tid": 537,
    "ts": 7429564854779.862, "dur": 81.652,
    "args": {
      "Python parent id": 4330, "Python module id": 187, "Python id": 4349, "finished": true, "Ev Idx": 11179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564854780.280, "dur": 80.984,
    "args": {
      "Python parent id": 4349, "Python id": 4350, "finished": true, "Ev Idx": 11180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564854780.465, "dur": 0.250,
    "args": {
      "Python parent id": 4350, "Python id": 4351, "finished": true, "Ev Idx": 11181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564854781.303, "dur": 79.789,
    "args": {
      "Python parent id": 4350, "Python id": 4352, "finished": true, "Ev Idx": 11182
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854781.555, "dur": 0.171,
    "args": {
      "Python parent id": 4352, "Python id": 4353, "finished": true, "Ev Idx": 11183
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854781.947, "dur": 0.115,
    "args": {
      "Python parent id": 4352, "Python id": 4354, "finished": true, "Ev Idx": 11184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564854782.198, "dur": 78.805,
    "args": {
      "Python parent id": 4352, "Python id": 4355, "finished": true, "Ev Idx": 11185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_188", "pid": 537, "tid": 537,
    "ts": 7429564854874.795, "dur": 45.437,
    "args": {
      "Python parent id": 4330, "Python module id": 188, "Python id": 4356, "finished": true, "Ev Idx": 11186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564854875.501, "dur": 44.543,
    "args": {
      "Python parent id": 4356, "Python id": 4357, "finished": true, "Ev Idx": 11187
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564854875.800, "dur": 0.482,
    "args": {
      "Python parent id": 4357, "Python id": 4358, "finished": true, "Ev Idx": 11188
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564854876.964, "dur": 42.963,
    "args": {
      "Python parent id": 4357, "Python id": 4359, "finished": true, "Ev Idx": 11189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854877.361, "dur": 0.241,
    "args": {
      "Python parent id": 4359, "Python id": 4360, "finished": true, "Ev Idx": 11190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854877.853, "dur": 0.171,
    "args": {
      "Python parent id": 4359, "Python id": 4361, "finished": true, "Ev Idx": 11191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564854878.162, "dur": 41.709,
    "args": {
      "Python parent id": 4359, "Python id": 4362, "finished": true, "Ev Idx": 11192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_27", "pid": 537, "tid": 537,
    "ts": 7429564854941.563, "dur": 1190.416,
    "args": {
      "Python parent id": 64, "Python module id": 27, "Python id": 4363, "finished": true, "Ev Idx": 11193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564854943.001, "dur": 1188.156,
    "args": {
      "Python parent id": 4363, "Python id": 4364, "finished": true, "Ev Idx": 11194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564854943.209, "dur": 0.372,
    "args": {
      "Python parent id": 4364, "Python id": 4365, "finished": true, "Ev Idx": 11195
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564854945.314, "dur": 1185.073,
    "args": {
      "Python parent id": 4364, "Python id": 4366, "finished": true, "Ev Idx": 11196
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564854945.781, "dur": 0.614,
    "args": {
      "Python parent id": 4366, "Python id": 4367, "finished": true, "Ev Idx": 11197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_54", "pid": 537, "tid": 537,
    "ts": 7429564854948.454, "dur": 82.081,
    "args": {
      "Python parent id": 4366, "Python module id": 54, "Python id": 4368, "finished": true, "Ev Idx": 11198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564854949.094, "dur": 80.978,
    "args": {
      "Python parent id": 4368, "Python id": 4369, "finished": true, "Ev Idx": 11199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564854949.236, "dur": 0.150,
    "args": {
      "Python parent id": 4369, "Python id": 4370, "finished": true, "Ev Idx": 11200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564854950.274, "dur": 78.978,
    "args": {
      "Python parent id": 4369, "Python id": 4371, "finished": true, "Ev Idx": 11201
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564854950.964, "dur": 1.923,
    "args": {
      "Python parent id": 4371, "Python id": 4372, "finished": true, "Ev Idx": 11202
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564854953.184, "dur": 12.492,
    "args": {
      "Python parent id": 4371, "Python id": 4373, "finished": true, "Ev Idx": 11203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564854966.043, "dur": 14.724,
    "args": {
      "Python parent id": 4371, "Python id": 4374, "finished": true, "Ev Idx": 11204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564854995.219, "dur": 10.324,
    "args": {
      "Python parent id": 4371, "Python id": 4375, "finished": true, "Ev Idx": 11205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855018.315, "dur": 0.217,
    "args": {
      "Python parent id": 4371, "Python id": 4376, "finished": true, "Ev Idx": 11206
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564855018.862, "dur": 1.295,
    "args": {
      "Python parent id": 4371, "Python id": 4377, "finished": true, "Ev Idx": 11207
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855031.216, "dur": 0.572,
    "args": {
      "Python parent id": 4366, "Python id": 4378, "finished": true, "Ev Idx": 11208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_27", "pid": 537, "tid": 537,
    "ts": 7429564855034.953, "dur": 693.919,
    "args": {
      "Python parent id": 4366, "Python module id": 27, "Python id": 4379, "finished": true, "Ev Idx": 11209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564855036.333, "dur": 691.825,
    "args": {
      "Python parent id": 4379, "Python id": 4380, "finished": true, "Ev Idx": 11210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564855036.583, "dur": 0.248,
    "args": {
      "Python parent id": 4380, "Python id": 4381, "finished": true, "Ev Idx": 11211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564855038.525, "dur": 688.862,
    "args": {
      "Python parent id": 4380, "Python id": 4382, "finished": true, "Ev Idx": 11212
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855042.264, "dur": 0.319,
    "args": {
      "Python parent id": 4382, "Python id": 4383, "finished": true, "Ev Idx": 11213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_189", "pid": 537, "tid": 537,
    "ts": 7429564855044.616, "dur": 40.843,
    "args": {
      "Python parent id": 4382, "Python module id": 189, "Python id": 4384, "finished": true, "Ev Idx": 11214
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564855045.270, "dur": 39.888,
    "args": {
      "Python parent id": 4384, "Python id": 4385, "finished": true, "Ev Idx": 11215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564855045.426, "dur": 0.305,
    "args": {
      "Python parent id": 4385, "Python id": 4386, "finished": true, "Ev Idx": 11216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564855046.441, "dur": 38.614,
    "args": {
      "Python parent id": 4385, "Python id": 4387, "finished": true, "Ev Idx": 11217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855046.759, "dur": 0.128,
    "args": {
      "Python parent id": 4387, "Python id": 4388, "finished": true, "Ev Idx": 11218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855047.105, "dur": 0.114,
    "args": {
      "Python parent id": 4387, "Python id": 4389, "finished": true, "Ev Idx": 11219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564855047.362, "dur": 37.616,
    "args": {
      "Python parent id": 4387, "Python id": 4390, "finished": true, "Ev Idx": 11220
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564855086.119, "dur": 3.690,
    "args": {
      "Python parent id": 4382, "Python id": 4391, "finished": true, "Ev Idx": 11221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564855090.594, "dur": 3.880,
    "args": {
      "Python parent id": 4382, "Python id": 4392, "finished": true, "Ev Idx": 11222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855095.259, "dur": 0.392,
    "args": {
      "Python parent id": 4382, "Python id": 4393, "finished": true, "Ev Idx": 11223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_190", "pid": 537, "tid": 537,
    "ts": 7429564855097.616, "dur": 44.094,
    "args": {
      "Python parent id": 4382, "Python module id": 190, "Python id": 4394, "finished": true, "Ev Idx": 11224
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564855098.134, "dur": 43.330,
    "args": {
      "Python parent id": 4394, "Python id": 4395, "finished": true, "Ev Idx": 11225
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564855098.319, "dur": 0.311,
    "args": {
      "Python parent id": 4395, "Python id": 4396, "finished": true, "Ev Idx": 11226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564855099.234, "dur": 42.088,
    "args": {
      "Python parent id": 4395, "Python id": 4397, "finished": true, "Ev Idx": 11227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855099.483, "dur": 0.124,
    "args": {
      "Python parent id": 4397, "Python id": 4398, "finished": true, "Ev Idx": 11228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855099.822, "dur": 0.141,
    "args": {
      "Python parent id": 4397, "Python id": 4399, "finished": true, "Ev Idx": 11229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564855100.067, "dur": 41.189,
    "args": {
      "Python parent id": 4397, "Python id": 4400, "finished": true, "Ev Idx": 11230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564855142.268, "dur": 2.613,
    "args": {
      "Python parent id": 4382, "Python id": 4401, "finished": true, "Ev Idx": 11231
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564855145.433, "dur": 2.923,
    "args": {
      "Python parent id": 4382, "Python id": 4402, "finished": true, "Ev Idx": 11232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855149.049, "dur": 0.469,
    "args": {
      "Python parent id": 4382, "Python id": 4403, "finished": true, "Ev Idx": 11233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_191", "pid": 537, "tid": 537,
    "ts": 7429564855151.641, "dur": 43.160,
    "args": {
      "Python parent id": 4382, "Python module id": 191, "Python id": 4404, "finished": true, "Ev Idx": 11234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564855152.212, "dur": 42.384,
    "args": {
      "Python parent id": 4404, "Python id": 4405, "finished": true, "Ev Idx": 11235
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564855152.376, "dur": 0.259,
    "args": {
      "Python parent id": 4405, "Python id": 4406, "finished": true, "Ev Idx": 11236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564855153.379, "dur": 41.080,
    "args": {
      "Python parent id": 4405, "Python id": 4407, "finished": true, "Ev Idx": 11237
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855153.595, "dur": 0.139,
    "args": {
      "Python parent id": 4407, "Python id": 4408, "finished": true, "Ev Idx": 11238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855153.911, "dur": 0.143,
    "args": {
      "Python parent id": 4407, "Python id": 4409, "finished": true, "Ev Idx": 11239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564855154.157, "dur": 40.241,
    "args": {
      "Python parent id": 4407, "Python id": 4410, "finished": true, "Ev Idx": 11240
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564855195.439, "dur": 2.477,
    "args": {
      "Python parent id": 4382, "Python id": 4411, "finished": true, "Ev Idx": 11241
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564855198.537, "dur": 2.474,
    "args": {
      "Python parent id": 4382, "Python id": 4412, "finished": true, "Ev Idx": 11242
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564855202.157, "dur": 188.414,
    "args": {
      "Python parent id": 4382, "Python id": 4413, "finished": true, "Ev Idx": 11243
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564855202.419, "dur": 3.519,
    "args": {
      "Python parent id": 4413, "Python id": 4414, "finished": true, "Ev Idx": 11244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564855206.301, "dur": 1.846,
    "args": {
      "Python parent id": 4413, "Python id": 4415, "finished": true, "Ev Idx": 11245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564855219.172, "dur": 89.283,
    "args": {
      "Python parent id": 4413, "Python id": 4416, "finished": true, "Ev Idx": 11246
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564855246.045, "dur": 61.103,
    "args": {
      "Python parent id": 4416, "Python id": 4417, "finished": true, "Ev Idx": 11247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564855340.458, "dur": 31.425,
    "args": {
      "Python parent id": 4413, "Python id": 4418, "finished": true, "Ev Idx": 11248
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564855359.067, "dur": 11.941,
    "args": {
      "Python parent id": 4418, "Python id": 4419, "finished": true, "Ev Idx": 11249
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564855394.978, "dur": 4.030,
    "args": {
      "Python parent id": 4382, "Python id": 4420, "finished": true, "Ev Idx": 11250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564855395.724, "dur": 0.321,
    "args": {
      "Python parent id": 4420, "Python id": 4421, "finished": true, "Ev Idx": 11251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564855396.366, "dur": 0.095,
    "args": {
      "Python parent id": 4420, "Python id": 4422, "finished": true, "Ev Idx": 11252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564855397.954, "dur": 0.257,
    "args": {
      "Python parent id": 4420, "Python id": 4423, "finished": true, "Ev Idx": 11253
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564855398.655, "dur": 0.116,
    "args": {
      "Python parent id": 4420, "Python id": 4424, "finished": true, "Ev Idx": 11254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564855400.399, "dur": 6.376,
    "args": {
      "Python parent id": 4382, "Python id": 4425, "finished": true, "Ev Idx": 11255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564855403.383, "dur": 3.249,
    "args": {
      "Python parent id": 4425, "Python id": 4426, "finished": true, "Ev Idx": 11256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564855403.615, "dur": 1.433,
    "args": {
      "Python parent id": 4426, "Python id": 4427, "finished": true, "Ev Idx": 11257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564855404.151, "dur": 0.743,
    "args": {
      "Python parent id": 4427, "Python id": 4428, "finished": true, "Ev Idx": 11258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564855405.318, "dur": 0.517,
    "args": {
      "Python parent id": 4426, "Python id": 4429, "finished": true, "Ev Idx": 11259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564855406.117, "dur": 0.441,
    "args": {
      "Python parent id": 4426, "Python id": 4430, "finished": true, "Ev Idx": 11260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564855409.634, "dur": 260.235,
    "args": {
      "Python parent id": 4382, "Python id": 4431, "finished": true, "Ev Idx": 11261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564855410.184, "dur": 46.049,
    "args": {
      "Python parent id": 4431, "Python id": 4432, "finished": true, "Ev Idx": 11262
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564855428.427, "dur": 5.111,
    "args": {
      "Python parent id": 4432, "Python id": 4433, "finished": true, "Ev Idx": 11263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564855434.479, "dur": 21.612,
    "args": {
      "Python parent id": 4432, "Python id": 4434, "finished": true, "Ev Idx": 11264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564855457.044, "dur": 31.639,
    "args": {
      "Python parent id": 4431, "Python id": 4435, "finished": true, "Ev Idx": 11265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564855470.598, "dur": 3.229,
    "args": {
      "Python parent id": 4435, "Python id": 4436, "finished": true, "Ev Idx": 11266
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564855474.418, "dur": 14.188,
    "args": {
      "Python parent id": 4435, "Python id": 4437, "finished": true, "Ev Idx": 11267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564855489.584, "dur": 3.025,
    "args": {
      "Python parent id": 4431, "Python id": 4438, "finished": true, "Ev Idx": 11268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564855492.899, "dur": 49.479,
    "args": {
      "Python parent id": 4431, "Python id": 4439, "finished": true, "Ev Idx": 11269
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564855575.501, "dur": 16.405,
    "args": {
      "Python parent id": 4431, "Python id": 4440, "finished": true, "Ev Idx": 11270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564855575.672, "dur": 0.259,
    "args": {
      "Python parent id": 4440, "Python id": 4441, "finished": true, "Ev Idx": 11271
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564855576.599, "dur": 15.155,
    "args": {
      "Python parent id": 4440, "Python id": 4442, "finished": true, "Ev Idx": 11272
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564855592.750, "dur": 1.758,
    "args": {
      "Python parent id": 4431, "Python id": 4443, "finished": true, "Ev Idx": 11273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564855595.936, "dur": 9.593,
    "args": {
      "Python parent id": 4431, "Python id": 4444, "finished": true, "Ev Idx": 11274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564855596.080, "dur": 0.133,
    "args": {
      "Python parent id": 4444, "Python id": 4445, "finished": true, "Ev Idx": 11275
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855601.191, "dur": 1.099,
    "args": {
      "Python parent id": 4444, "Python id": 4446, "finished": true, "Ev Idx": 11276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564855601.714, "dur": 0.507,
    "args": {
      "Python parent id": 4446, "Python id": 4447, "finished": true, "Ev Idx": 11277
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564855602.675, "dur": 2.776,
    "args": {
      "Python parent id": 4444, "Python id": 4448, "finished": true, "Ev Idx": 11278
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564855606.279, "dur": 44.345,
    "args": {
      "Python parent id": 4431, "Python id": 4449, "finished": true, "Ev Idx": 11279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564855651.150, "dur": 2.954,
    "args": {
      "Python parent id": 4431, "Python id": 4450, "finished": true, "Ev Idx": 11280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564855654.721, "dur": 14.189,
    "args": {
      "Python parent id": 4431, "Python id": 4451, "finished": true, "Ev Idx": 11281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564855672.523, "dur": 3.209,
    "args": {
      "Python parent id": 4382, "Python id": 4452, "finished": true, "Ev Idx": 11282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564855676.415, "dur": 0.288,
    "args": {
      "Python parent id": 4382, "Python id": 4453, "finished": true, "Ev Idx": 11283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855677.762, "dur": 0.799,
    "args": {
      "Python parent id": 4382, "Python id": 4454, "finished": true, "Ev Idx": 11284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_192", "pid": 537, "tid": 537,
    "ts": 7429564855682.755, "dur": 44.110,
    "args": {
      "Python parent id": 4382, "Python module id": 192, "Python id": 4455, "finished": true, "Ev Idx": 11285
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564855683.692, "dur": 42.895,
    "args": {
      "Python parent id": 4455, "Python id": 4456, "finished": true, "Ev Idx": 11286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564855684.094, "dur": 1.047,
    "args": {
      "Python parent id": 4456, "Python id": 4457, "finished": true, "Ev Idx": 11287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564855686.392, "dur": 39.996,
    "args": {
      "Python parent id": 4456, "Python id": 4458, "finished": true, "Ev Idx": 11288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855686.816, "dur": 0.237,
    "args": {
      "Python parent id": 4458, "Python id": 4459, "finished": true, "Ev Idx": 11289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855687.347, "dur": 0.186,
    "args": {
      "Python parent id": 4458, "Python id": 4460, "finished": true, "Ev Idx": 11290
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564855687.695, "dur": 38.597,
    "args": {
      "Python parent id": 4458, "Python id": 4461, "finished": true, "Ev Idx": 11291
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855742.164, "dur": 0.793,
    "args": {
      "Python parent id": 4366, "Python id": 4462, "finished": true, "Ev Idx": 11292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_55", "pid": 537, "tid": 537,
    "ts": 7429564855745.503, "dur": 83.298,
    "args": {
      "Python parent id": 4366, "Python module id": 55, "Python id": 4463, "finished": true, "Ev Idx": 11293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564855746.198, "dur": 82.147,
    "args": {
      "Python parent id": 4463, "Python id": 4464, "finished": true, "Ev Idx": 11294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564855746.402, "dur": 0.435,
    "args": {
      "Python parent id": 4464, "Python id": 4465, "finished": true, "Ev Idx": 11295
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564855747.570, "dur": 79.973,
    "args": {
      "Python parent id": 4464, "Python id": 4466, "finished": true, "Ev Idx": 11296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564855748.108, "dur": 1.592,
    "args": {
      "Python parent id": 4466, "Python id": 4467, "finished": true, "Ev Idx": 11297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564855750.034, "dur": 13.469,
    "args": {
      "Python parent id": 4466, "Python id": 4468, "finished": true, "Ev Idx": 11298
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564855763.859, "dur": 15.253,
    "args": {
      "Python parent id": 4466, "Python id": 4469, "finished": true, "Ev Idx": 11299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564855790.903, "dur": 11.870,
    "args": {
      "Python parent id": 4466, "Python id": 4470, "finished": true, "Ev Idx": 11300
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855815.607, "dur": 0.208,
    "args": {
      "Python parent id": 4466, "Python id": 4471, "finished": true, "Ev Idx": 11301
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564855816.179, "dur": 1.249,
    "args": {
      "Python parent id": 4466, "Python id": 4472, "finished": true, "Ev Idx": 11302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855829.527, "dur": 0.415,
    "args": {
      "Python parent id": 4366, "Python id": 4473, "finished": true, "Ev Idx": 11303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_27", "pid": 537, "tid": 537,
    "ts": 7429564855831.502, "dur": 287.538,
    "args": {
      "Python parent id": 4366, "Python module id": 27, "Python id": 4474, "finished": true, "Ev Idx": 11304
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564855832.063, "dur": 286.550,
    "args": {
      "Python parent id": 4474, "Python id": 4475, "finished": true, "Ev Idx": 11305
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564855832.276, "dur": 0.359,
    "args": {
      "Python parent id": 4475, "Python id": 4476, "finished": true, "Ev Idx": 11306
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564855833.523, "dur": 284.993,
    "args": {
      "Python parent id": 4475, "Python id": 4477, "finished": true, "Ev Idx": 11307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855833.761, "dur": 0.500,
    "args": {
      "Python parent id": 4477, "Python id": 4478, "finished": true, "Ev Idx": 11308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855834.589, "dur": 0.297,
    "args": {
      "Python parent id": 4477, "Python id": 4479, "finished": true, "Ev Idx": 11309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855835.109, "dur": 0.223,
    "args": {
      "Python parent id": 4477, "Python id": 4480, "finished": true, "Ev Idx": 11310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_193", "pid": 537, "tid": 537,
    "ts": 7429564855837.556, "dur": 53.902,
    "args": {
      "Python parent id": 4477, "Python module id": 193, "Python id": 4481, "finished": true, "Ev Idx": 11311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564855838.139, "dur": 53.039,
    "args": {
      "Python parent id": 4481, "Python id": 4482, "finished": true, "Ev Idx": 11312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564855838.301, "dur": 0.179,
    "args": {
      "Python parent id": 4482, "Python id": 4483, "finished": true, "Ev Idx": 11313
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564855839.168, "dur": 51.855,
    "args": {
      "Python parent id": 4482, "Python id": 4484, "finished": true, "Ev Idx": 11314
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855839.438, "dur": 0.187,
    "args": {
      "Python parent id": 4484, "Python id": 4485, "finished": true, "Ev Idx": 11315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855839.850, "dur": 0.159,
    "args": {
      "Python parent id": 4484, "Python id": 4486, "finished": true, "Ev Idx": 11316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564855840.143, "dur": 50.760,
    "args": {
      "Python parent id": 4484, "Python id": 4487, "finished": true, "Ev Idx": 11317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_27", "pid": 537, "tid": 537,
    "ts": 7429564855893.690, "dur": 67.934,
    "args": {
      "Python parent id": 4477, "Python module id": 27, "Python id": 4488, "finished": true, "Ev Idx": 11318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564855894.233, "dur": 67.087,
    "args": {
      "Python parent id": 4488, "Python id": 4489, "finished": true, "Ev Idx": 11319
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564855894.489, "dur": 0.355,
    "args": {
      "Python parent id": 4489, "Python id": 4490, "finished": true, "Ev Idx": 11320
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564855895.840, "dur": 65.309,
    "args": {
      "Python parent id": 4489, "Python id": 4491, "finished": true, "Ev Idx": 11321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564855896.724, "dur": 64.244,
    "args": {
      "Python parent id": 4491, "Python id": 4492, "finished": true, "Ev Idx": 11322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564855896.869, "dur": 0.212,
    "args": {
      "Python parent id": 4492, "Python id": 4493, "finished": true, "Ev Idx": 11323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564855897.940, "dur": 62.920,
    "args": {
      "Python parent id": 4492, "Python id": 4494, "finished": true, "Ev Idx": 11324
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855962.681, "dur": 0.423,
    "args": {
      "Python parent id": 4477, "Python id": 4495, "finished": true, "Ev Idx": 11325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_194", "pid": 537, "tid": 537,
    "ts": 7429564855965.410, "dur": 53.151,
    "args": {
      "Python parent id": 4477, "Python module id": 194, "Python id": 4496, "finished": true, "Ev Idx": 11326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564855966.331, "dur": 52.029,
    "args": {
      "Python parent id": 4496, "Python id": 4497, "finished": true, "Ev Idx": 11327
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564855966.586, "dur": 0.414,
    "args": {
      "Python parent id": 4497, "Python id": 4498, "finished": true, "Ev Idx": 11328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564855967.744, "dur": 50.521,
    "args": {
      "Python parent id": 4497, "Python id": 4499, "finished": true, "Ev Idx": 11329
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855968.052, "dur": 0.231,
    "args": {
      "Python parent id": 4499, "Python id": 4500, "finished": true, "Ev Idx": 11330
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564855968.470, "dur": 0.117,
    "args": {
      "Python parent id": 4499, "Python id": 4501, "finished": true, "Ev Idx": 11331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564855968.855, "dur": 49.311,
    "args": {
      "Python parent id": 4499, "Python id": 4502, "finished": true, "Ev Idx": 11332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_195", "pid": 537, "tid": 537,
    "ts": 7429564856065.480, "dur": 52.738,
    "args": {
      "Python parent id": 4477, "Python module id": 195, "Python id": 4503, "finished": true, "Ev Idx": 11333
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564856066.241, "dur": 51.729,
    "args": {
      "Python parent id": 4503, "Python id": 4504, "finished": true, "Ev Idx": 11334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564856066.538, "dur": 0.429,
    "args": {
      "Python parent id": 4504, "Python id": 4505, "finished": true, "Ev Idx": 11335
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564856067.616, "dur": 50.229,
    "args": {
      "Python parent id": 4504, "Python id": 4506, "finished": true, "Ev Idx": 11336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856067.962, "dur": 0.312,
    "args": {
      "Python parent id": 4506, "Python id": 4507, "finished": true, "Ev Idx": 11337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856068.536, "dur": 0.155,
    "args": {
      "Python parent id": 4506, "Python id": 4508, "finished": true, "Ev Idx": 11338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564856068.838, "dur": 48.953,
    "args": {
      "Python parent id": 4506, "Python id": 4509, "finished": true, "Ev Idx": 11339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_28", "pid": 537, "tid": 537,
    "ts": 7429564856139.052, "dur": 1171.460,
    "args": {
      "Python parent id": 64, "Python module id": 28, "Python id": 4510, "finished": true, "Ev Idx": 11340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564856140.904, "dur": 1168.783,
    "args": {
      "Python parent id": 4510, "Python id": 4511, "finished": true, "Ev Idx": 11341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564856141.222, "dur": 0.313,
    "args": {
      "Python parent id": 4511, "Python id": 4512, "finished": true, "Ev Idx": 11342
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564856143.258, "dur": 1165.672,
    "args": {
      "Python parent id": 4511, "Python id": 4513, "finished": true, "Ev Idx": 11343
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856143.809, "dur": 0.398,
    "args": {
      "Python parent id": 4513, "Python id": 4514, "finished": true, "Ev Idx": 11344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_56", "pid": 537, "tid": 537,
    "ts": 7429564856146.311, "dur": 80.537,
    "args": {
      "Python parent id": 4513, "Python module id": 56, "Python id": 4515, "finished": true, "Ev Idx": 11345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564856146.931, "dur": 79.545,
    "args": {
      "Python parent id": 4515, "Python id": 4516, "finished": true, "Ev Idx": 11346
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564856147.075, "dur": 0.176,
    "args": {
      "Python parent id": 4516, "Python id": 4517, "finished": true, "Ev Idx": 11347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564856147.973, "dur": 77.627,
    "args": {
      "Python parent id": 4516, "Python id": 4518, "finished": true, "Ev Idx": 11348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564856148.611, "dur": 2.141,
    "args": {
      "Python parent id": 4518, "Python id": 4519, "finished": true, "Ev Idx": 11349
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564856151.102, "dur": 14.086,
    "args": {
      "Python parent id": 4518, "Python id": 4520, "finished": true, "Ev Idx": 11350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564856165.551, "dur": 14.617,
    "args": {
      "Python parent id": 4518, "Python id": 4521, "finished": true, "Ev Idx": 11351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564856192.191, "dur": 10.326,
    "args": {
      "Python parent id": 4518, "Python id": 4522, "finished": true, "Ev Idx": 11352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856212.705, "dur": 0.252,
    "args": {
      "Python parent id": 4518, "Python id": 4523, "finished": true, "Ev Idx": 11353
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564856213.424, "dur": 1.163,
    "args": {
      "Python parent id": 4518, "Python id": 4524, "finished": true, "Ev Idx": 11354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856227.463, "dur": 0.376,
    "args": {
      "Python parent id": 4513, "Python id": 4525, "finished": true, "Ev Idx": 11355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_28", "pid": 537, "tid": 537,
    "ts": 7429564856230.935, "dur": 649.475,
    "args": {
      "Python parent id": 4513, "Python module id": 28, "Python id": 4526, "finished": true, "Ev Idx": 11356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564856232.403, "dur": 647.345,
    "args": {
      "Python parent id": 4526, "Python id": 4527, "finished": true, "Ev Idx": 11357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564856232.600, "dur": 0.273,
    "args": {
      "Python parent id": 4527, "Python id": 4528, "finished": true, "Ev Idx": 11358
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564856234.527, "dur": 644.485,
    "args": {
      "Python parent id": 4527, "Python id": 4529, "finished": true, "Ev Idx": 11359
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856238.122, "dur": 0.345,
    "args": {
      "Python parent id": 4529, "Python id": 4530, "finished": true, "Ev Idx": 11360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_196", "pid": 537, "tid": 537,
    "ts": 7429564856240.185, "dur": 43.229,
    "args": {
      "Python parent id": 4529, "Python module id": 196, "Python id": 4531, "finished": true, "Ev Idx": 11361
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564856240.784, "dur": 42.310,
    "args": {
      "Python parent id": 4531, "Python id": 4532, "finished": true, "Ev Idx": 11362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564856240.935, "dur": 0.210,
    "args": {
      "Python parent id": 4532, "Python id": 4533, "finished": true, "Ev Idx": 11363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564856241.985, "dur": 40.856,
    "args": {
      "Python parent id": 4532, "Python id": 4534, "finished": true, "Ev Idx": 11364
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856242.282, "dur": 0.117,
    "args": {
      "Python parent id": 4534, "Python id": 4535, "finished": true, "Ev Idx": 11365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856242.639, "dur": 0.128,
    "args": {
      "Python parent id": 4534, "Python id": 4536, "finished": true, "Ev Idx": 11366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564856242.920, "dur": 39.808,
    "args": {
      "Python parent id": 4534, "Python id": 4537, "finished": true, "Ev Idx": 11367
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564856284.035, "dur": 3.368,
    "args": {
      "Python parent id": 4529, "Python id": 4538, "finished": true, "Ev Idx": 11368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564856287.979, "dur": 3.993,
    "args": {
      "Python parent id": 4529, "Python id": 4539, "finished": true, "Ev Idx": 11369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856292.877, "dur": 0.353,
    "args": {
      "Python parent id": 4529, "Python id": 4540, "finished": true, "Ev Idx": 11370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_197", "pid": 537, "tid": 537,
    "ts": 7429564856295.294, "dur": 44.763,
    "args": {
      "Python parent id": 4529, "Python module id": 197, "Python id": 4541, "finished": true, "Ev Idx": 11371
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564856295.799, "dur": 44.064,
    "args": {
      "Python parent id": 4541, "Python id": 4542, "finished": true, "Ev Idx": 11372
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564856296.089, "dur": 0.250,
    "args": {
      "Python parent id": 4542, "Python id": 4543, "finished": true, "Ev Idx": 11373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564856296.979, "dur": 42.796,
    "args": {
      "Python parent id": 4542, "Python id": 4544, "finished": true, "Ev Idx": 11374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856297.209, "dur": 0.122,
    "args": {
      "Python parent id": 4544, "Python id": 4545, "finished": true, "Ev Idx": 11375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856297.511, "dur": 0.121,
    "args": {
      "Python parent id": 4544, "Python id": 4546, "finished": true, "Ev Idx": 11376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564856297.756, "dur": 41.949,
    "args": {
      "Python parent id": 4544, "Python id": 4547, "finished": true, "Ev Idx": 11377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564856340.530, "dur": 2.360,
    "args": {
      "Python parent id": 4529, "Python id": 4548, "finished": true, "Ev Idx": 11378
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564856343.457, "dur": 2.762,
    "args": {
      "Python parent id": 4529, "Python id": 4549, "finished": true, "Ev Idx": 11379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856346.962, "dur": 0.484,
    "args": {
      "Python parent id": 4529, "Python id": 4550, "finished": true, "Ev Idx": 11380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_198", "pid": 537, "tid": 537,
    "ts": 7429564856349.849, "dur": 39.950,
    "args": {
      "Python parent id": 4529, "Python module id": 198, "Python id": 4551, "finished": true, "Ev Idx": 11381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564856350.410, "dur": 39.124,
    "args": {
      "Python parent id": 4551, "Python id": 4552, "finished": true, "Ev Idx": 11382
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564856350.668, "dur": 0.262,
    "args": {
      "Python parent id": 4552, "Python id": 4553, "finished": true, "Ev Idx": 11383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564856351.524, "dur": 37.932,
    "args": {
      "Python parent id": 4552, "Python id": 4554, "finished": true, "Ev Idx": 11384
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856351.747, "dur": 0.179,
    "args": {
      "Python parent id": 4554, "Python id": 4555, "finished": true, "Ev Idx": 11385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856352.142, "dur": 0.119,
    "args": {
      "Python parent id": 4554, "Python id": 4556, "finished": true, "Ev Idx": 11386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564856352.399, "dur": 36.952,
    "args": {
      "Python parent id": 4554, "Python id": 4557, "finished": true, "Ev Idx": 11387
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564856390.370, "dur": 2.396,
    "args": {
      "Python parent id": 4529, "Python id": 4558, "finished": true, "Ev Idx": 11388
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564856393.256, "dur": 2.432,
    "args": {
      "Python parent id": 4529, "Python id": 4559, "finished": true, "Ev Idx": 11389
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564856396.914, "dur": 149.528,
    "args": {
      "Python parent id": 4529, "Python id": 4560, "finished": true, "Ev Idx": 11390
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564856397.179, "dur": 3.662,
    "args": {
      "Python parent id": 4560, "Python id": 4561, "finished": true, "Ev Idx": 11391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564856401.170, "dur": 1.827,
    "args": {
      "Python parent id": 4560, "Python id": 4562, "finished": true, "Ev Idx": 11392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564856415.057, "dur": 51.193,
    "args": {
      "Python parent id": 4560, "Python id": 4563, "finished": true, "Ev Idx": 11393
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564856444.533, "dur": 20.604,
    "args": {
      "Python parent id": 4563, "Python id": 4564, "finished": true, "Ev Idx": 11394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564856497.952, "dur": 30.134,
    "args": {
      "Python parent id": 4560, "Python id": 4565, "finished": true, "Ev Idx": 11395
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564856515.549, "dur": 11.667,
    "args": {
      "Python parent id": 4565, "Python id": 4566, "finished": true, "Ev Idx": 11396
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564856550.777, "dur": 3.951,
    "args": {
      "Python parent id": 4529, "Python id": 4567, "finished": true, "Ev Idx": 11397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564856551.636, "dur": 0.369,
    "args": {
      "Python parent id": 4567, "Python id": 4568, "finished": true, "Ev Idx": 11398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564856552.342, "dur": 0.084,
    "args": {
      "Python parent id": 4567, "Python id": 4569, "finished": true, "Ev Idx": 11399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564856553.707, "dur": 0.263,
    "args": {
      "Python parent id": 4567, "Python id": 4570, "finished": true, "Ev Idx": 11400
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564856554.379, "dur": 0.115,
    "args": {
      "Python parent id": 4567, "Python id": 4571, "finished": true, "Ev Idx": 11401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564856556.065, "dur": 6.939,
    "args": {
      "Python parent id": 4529, "Python id": 4572, "finished": true, "Ev Idx": 11402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564856559.285, "dur": 3.559,
    "args": {
      "Python parent id": 4572, "Python id": 4573, "finished": true, "Ev Idx": 11403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564856559.598, "dur": 1.616,
    "args": {
      "Python parent id": 4573, "Python id": 4574, "finished": true, "Ev Idx": 11404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564856560.144, "dur": 0.902,
    "args": {
      "Python parent id": 4574, "Python id": 4575, "finished": true, "Ev Idx": 11405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564856561.501, "dur": 0.484,
    "args": {
      "Python parent id": 4573, "Python id": 4576, "finished": true, "Ev Idx": 11406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564856562.321, "dur": 0.433,
    "args": {
      "Python parent id": 4573, "Python id": 4577, "finished": true, "Ev Idx": 11407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564856565.795, "dur": 259.396,
    "args": {
      "Python parent id": 4529, "Python id": 4578, "finished": true, "Ev Idx": 11408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564856566.335, "dur": 47.582,
    "args": {
      "Python parent id": 4578, "Python id": 4579, "finished": true, "Ev Idx": 11409
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564856584.803, "dur": 6.119,
    "args": {
      "Python parent id": 4579, "Python id": 4580, "finished": true, "Ev Idx": 11410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564856591.806, "dur": 21.968,
    "args": {
      "Python parent id": 4579, "Python id": 4581, "finished": true, "Ev Idx": 11411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564856614.841, "dur": 32.411,
    "args": {
      "Python parent id": 4578, "Python id": 4582, "finished": true, "Ev Idx": 11412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564856632.167, "dur": 2.587,
    "args": {
      "Python parent id": 4582, "Python id": 4583, "finished": true, "Ev Idx": 11413
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564856635.277, "dur": 11.912,
    "args": {
      "Python parent id": 4582, "Python id": 4584, "finished": true, "Ev Idx": 11414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564856648.127, "dur": 3.989,
    "args": {
      "Python parent id": 4578, "Python id": 4585, "finished": true, "Ev Idx": 11415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564856652.417, "dur": 46.468,
    "args": {
      "Python parent id": 4578, "Python id": 4586, "finished": true, "Ev Idx": 11416
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564856732.917, "dur": 16.250,
    "args": {
      "Python parent id": 4578, "Python id": 4587, "finished": true, "Ev Idx": 11417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564856733.218, "dur": 0.224,
    "args": {
      "Python parent id": 4587, "Python id": 4588, "finished": true, "Ev Idx": 11418
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564856734.201, "dur": 14.764,
    "args": {
      "Python parent id": 4587, "Python id": 4589, "finished": true, "Ev Idx": 11419
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564856750.236, "dur": 1.738,
    "args": {
      "Python parent id": 4578, "Python id": 4590, "finished": true, "Ev Idx": 11420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564856753.540, "dur": 11.141,
    "args": {
      "Python parent id": 4578, "Python id": 4591, "finished": true, "Ev Idx": 11421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564856753.614, "dur": 0.112,
    "args": {
      "Python parent id": 4591, "Python id": 4592, "finished": true, "Ev Idx": 11422
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856760.510, "dur": 1.044,
    "args": {
      "Python parent id": 4591, "Python id": 4593, "finished": true, "Ev Idx": 11423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564856761.038, "dur": 0.411,
    "args": {
      "Python parent id": 4593, "Python id": 4594, "finished": true, "Ev Idx": 11424
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564856761.971, "dur": 2.621,
    "args": {
      "Python parent id": 4591, "Python id": 4595, "finished": true, "Ev Idx": 11425
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564856765.350, "dur": 40.971,
    "args": {
      "Python parent id": 4578, "Python id": 4596, "finished": true, "Ev Idx": 11426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564856806.937, "dur": 3.156,
    "args": {
      "Python parent id": 4578, "Python id": 4597, "finished": true, "Ev Idx": 11427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564856810.777, "dur": 13.440,
    "args": {
      "Python parent id": 4578, "Python id": 4598, "finished": true, "Ev Idx": 11428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564856828.085, "dur": 3.163,
    "args": {
      "Python parent id": 4529, "Python id": 4599, "finished": true, "Ev Idx": 11429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564856831.899, "dur": 0.226,
    "args": {
      "Python parent id": 4529, "Python id": 4600, "finished": true, "Ev Idx": 11430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856833.155, "dur": 0.826,
    "args": {
      "Python parent id": 4529, "Python id": 4601, "finished": true, "Ev Idx": 11431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_199", "pid": 537, "tid": 537,
    "ts": 7429564856837.996, "dur": 40.488,
    "args": {
      "Python parent id": 4529, "Python module id": 199, "Python id": 4602, "finished": true, "Ev Idx": 11432
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564856838.768, "dur": 39.338,
    "args": {
      "Python parent id": 4602, "Python id": 4603, "finished": true, "Ev Idx": 11433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564856839.155, "dur": 0.782,
    "args": {
      "Python parent id": 4603, "Python id": 4604, "finished": true, "Ev Idx": 11434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564856840.973, "dur": 36.893,
    "args": {
      "Python parent id": 4603, "Python id": 4605, "finished": true, "Ev Idx": 11435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856841.371, "dur": 0.246,
    "args": {
      "Python parent id": 4605, "Python id": 4606, "finished": true, "Ev Idx": 11436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856841.864, "dur": 0.137,
    "args": {
      "Python parent id": 4605, "Python id": 4607, "finished": true, "Ev Idx": 11437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564856842.204, "dur": 35.537,
    "args": {
      "Python parent id": 4605, "Python id": 4608, "finished": true, "Ev Idx": 11438
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856893.169, "dur": 0.560,
    "args": {
      "Python parent id": 4513, "Python id": 4609, "finished": true, "Ev Idx": 11439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_57", "pid": 537, "tid": 537,
    "ts": 7429564856896.218, "dur": 165.956,
    "args": {
      "Python parent id": 4513, "Python module id": 57, "Python id": 4610, "finished": true, "Ev Idx": 11440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564856897.224, "dur": 164.339,
    "args": {
      "Python parent id": 4610, "Python id": 4611, "finished": true, "Ev Idx": 11441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564856897.524, "dur": 0.409,
    "args": {
      "Python parent id": 4611, "Python id": 4612, "finished": true, "Ev Idx": 11442
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564856898.831, "dur": 161.783,
    "args": {
      "Python parent id": 4611, "Python id": 4613, "finished": true, "Ev Idx": 11443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564856899.380, "dur": 1.554,
    "args": {
      "Python parent id": 4613, "Python id": 4614, "finished": true, "Ev Idx": 11444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564856901.230, "dur": 12.584,
    "args": {
      "Python parent id": 4613, "Python id": 4615, "finished": true, "Ev Idx": 11445
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564856914.194, "dur": 13.793,
    "args": {
      "Python parent id": 4613, "Python id": 4616, "finished": true, "Ev Idx": 11446
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564856939.995, "dur": 11.040,
    "args": {
      "Python parent id": 4613, "Python id": 4617, "finished": true, "Ev Idx": 11447
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564856960.390, "dur": 0.307,
    "args": {
      "Python parent id": 4613, "Python id": 4618, "finished": true, "Ev Idx": 11448
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564856961.122, "dur": 1.130,
    "args": {
      "Python parent id": 4613, "Python id": 4619, "finished": true, "Ev Idx": 11449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857063.029, "dur": 1.851,
    "args": {
      "Python parent id": 4513, "Python id": 4620, "finished": true, "Ev Idx": 11450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_28", "pid": 537, "tid": 537,
    "ts": 7429564857067.068, "dur": 230.306,
    "args": {
      "Python parent id": 4513, "Python module id": 28, "Python id": 4621, "finished": true, "Ev Idx": 11451
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857067.701, "dur": 229.388,
    "args": {
      "Python parent id": 4621, "Python id": 4622, "finished": true, "Ev Idx": 11452
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857068.034, "dur": 0.596,
    "args": {
      "Python parent id": 4622, "Python id": 4623, "finished": true, "Ev Idx": 11453
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564857069.954, "dur": 227.032,
    "args": {
      "Python parent id": 4622, "Python id": 4624, "finished": true, "Ev Idx": 11454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857070.229, "dur": 0.354,
    "args": {
      "Python parent id": 4624, "Python id": 4625, "finished": true, "Ev Idx": 11455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857070.958, "dur": 0.319,
    "args": {
      "Python parent id": 4624, "Python id": 4626, "finished": true, "Ev Idx": 11456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857071.502, "dur": 0.252,
    "args": {
      "Python parent id": 4624, "Python id": 4627, "finished": true, "Ev Idx": 11457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_200", "pid": 537, "tid": 537,
    "ts": 7429564857073.809, "dur": 57.372,
    "args": {
      "Python parent id": 4624, "Python module id": 200, "Python id": 4628, "finished": true, "Ev Idx": 11458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857074.440, "dur": 56.381,
    "args": {
      "Python parent id": 4628, "Python id": 4629, "finished": true, "Ev Idx": 11459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857074.580, "dur": 0.184,
    "args": {
      "Python parent id": 4629, "Python id": 4630, "finished": true, "Ev Idx": 11460
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564857075.488, "dur": 55.231,
    "args": {
      "Python parent id": 4629, "Python id": 4631, "finished": true, "Ev Idx": 11461
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857075.750, "dur": 0.152,
    "args": {
      "Python parent id": 4631, "Python id": 4632, "finished": true, "Ev Idx": 11462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857076.144, "dur": 0.143,
    "args": {
      "Python parent id": 4631, "Python id": 4633, "finished": true, "Ev Idx": 11463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564857076.467, "dur": 54.137,
    "args": {
      "Python parent id": 4631, "Python id": 4634, "finished": true, "Ev Idx": 11464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_28", "pid": 537, "tid": 537,
    "ts": 7429564857133.121, "dur": 15.960,
    "args": {
      "Python parent id": 4624, "Python module id": 28, "Python id": 4635, "finished": true, "Ev Idx": 11465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857133.523, "dur": 15.376,
    "args": {
      "Python parent id": 4635, "Python id": 4636, "finished": true, "Ev Idx": 11466
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857133.712, "dur": 0.274,
    "args": {
      "Python parent id": 4636, "Python id": 4637, "finished": true, "Ev Idx": 11467
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564857134.989, "dur": 13.814,
    "args": {
      "Python parent id": 4636, "Python id": 4638, "finished": true, "Ev Idx": 11468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564857135.978, "dur": 12.677,
    "args": {
      "Python parent id": 4638, "Python id": 4639, "finished": true, "Ev Idx": 11469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564857136.143, "dur": 0.246,
    "args": {
      "Python parent id": 4639, "Python id": 4640, "finished": true, "Ev Idx": 11470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564857137.332, "dur": 11.232,
    "args": {
      "Python parent id": 4639, "Python id": 4641, "finished": true, "Ev Idx": 11471
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857149.789, "dur": 0.372,
    "args": {
      "Python parent id": 4624, "Python id": 4642, "finished": true, "Ev Idx": 11472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_201", "pid": 537, "tid": 537,
    "ts": 7429564857152.089, "dur": 85.380,
    "args": {
      "Python parent id": 4624, "Python module id": 201, "Python id": 4643, "finished": true, "Ev Idx": 11473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857152.719, "dur": 84.424,
    "args": {
      "Python parent id": 4643, "Python id": 4644, "finished": true, "Ev Idx": 11474
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857152.881, "dur": 0.199,
    "args": {
      "Python parent id": 4644, "Python id": 4645, "finished": true, "Ev Idx": 11475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564857153.657, "dur": 83.360,
    "args": {
      "Python parent id": 4644, "Python id": 4646, "finished": true, "Ev Idx": 11476
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857153.913, "dur": 0.119,
    "args": {
      "Python parent id": 4646, "Python id": 4647, "finished": true, "Ev Idx": 11477
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857154.186, "dur": 0.114,
    "args": {
      "Python parent id": 4646, "Python id": 4648, "finished": true, "Ev Idx": 11478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564857154.410, "dur": 82.470,
    "args": {
      "Python parent id": 4646, "Python id": 4649, "finished": true, "Ev Idx": 11479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_202", "pid": 537, "tid": 537,
    "ts": 7429564857252.622, "dur": 44.022,
    "args": {
      "Python parent id": 4624, "Python module id": 202, "Python id": 4650, "finished": true, "Ev Idx": 11480
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857253.319, "dur": 43.089,
    "args": {
      "Python parent id": 4650, "Python id": 4651, "finished": true, "Ev Idx": 11481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857253.622, "dur": 0.510,
    "args": {
      "Python parent id": 4651, "Python id": 4652, "finished": true, "Ev Idx": 11482
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564857254.769, "dur": 41.502,
    "args": {
      "Python parent id": 4651, "Python id": 4653, "finished": true, "Ev Idx": 11483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857255.072, "dur": 0.195,
    "args": {
      "Python parent id": 4653, "Python id": 4654, "finished": true, "Ev Idx": 11484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857255.497, "dur": 0.133,
    "args": {
      "Python parent id": 4653, "Python id": 4655, "finished": true, "Ev Idx": 11485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564857255.746, "dur": 40.484,
    "args": {
      "Python parent id": 4653, "Python id": 4656, "finished": true, "Ev Idx": 11486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_29", "pid": 537, "tid": 537,
    "ts": 7429564857317.180, "dur": 1243.007,
    "args": {
      "Python parent id": 64, "Python module id": 29, "Python id": 4657, "finished": true, "Ev Idx": 11487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857318.647, "dur": 1240.660,
    "args": {
      "Python parent id": 4657, "Python id": 4658, "finished": true, "Ev Idx": 11488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857318.860, "dur": 0.399,
    "args": {
      "Python parent id": 4658, "Python id": 4659, "finished": true, "Ev Idx": 11489
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564857321.026, "dur": 1237.476,
    "args": {
      "Python parent id": 4658, "Python id": 4660, "finished": true, "Ev Idx": 11490
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857321.526, "dur": 0.445,
    "args": {
      "Python parent id": 4660, "Python id": 4661, "finished": true, "Ev Idx": 11491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_58", "pid": 537, "tid": 537,
    "ts": 7429564857324.003, "dur": 78.201,
    "args": {
      "Python parent id": 4660, "Python module id": 58, "Python id": 4662, "finished": true, "Ev Idx": 11492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857324.537, "dur": 77.389,
    "args": {
      "Python parent id": 4662, "Python id": 4663, "finished": true, "Ev Idx": 11493
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857324.680, "dur": 0.171,
    "args": {
      "Python parent id": 4663, "Python id": 4664, "finished": true, "Ev Idx": 11494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564857325.785, "dur": 75.342,
    "args": {
      "Python parent id": 4663, "Python id": 4665, "finished": true, "Ev Idx": 11495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564857326.397, "dur": 2.023,
    "args": {
      "Python parent id": 4665, "Python id": 4666, "finished": true, "Ev Idx": 11496
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564857328.722, "dur": 12.429,
    "args": {
      "Python parent id": 4665, "Python id": 4667, "finished": true, "Ev Idx": 11497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564857341.498, "dur": 15.070,
    "args": {
      "Python parent id": 4665, "Python id": 4668, "finished": true, "Ev Idx": 11498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564857370.125, "dur": 10.656,
    "args": {
      "Python parent id": 4665, "Python id": 4669, "finished": true, "Ev Idx": 11499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857390.041, "dur": 0.206,
    "args": {
      "Python parent id": 4665, "Python id": 4670, "finished": true, "Ev Idx": 11500
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564857390.786, "dur": 1.111,
    "args": {
      "Python parent id": 4665, "Python id": 4671, "finished": true, "Ev Idx": 11501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857402.825, "dur": 0.292,
    "args": {
      "Python parent id": 4660, "Python id": 4672, "finished": true, "Ev Idx": 11502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_29", "pid": 537, "tid": 537,
    "ts": 7429564857405.752, "dur": 730.581,
    "args": {
      "Python parent id": 4660, "Python module id": 29, "Python id": 4673, "finished": true, "Ev Idx": 11503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857407.177, "dur": 728.369,
    "args": {
      "Python parent id": 4673, "Python id": 4674, "finished": true, "Ev Idx": 11504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857407.400, "dur": 0.281,
    "args": {
      "Python parent id": 4674, "Python id": 4675, "finished": true, "Ev Idx": 11505
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564857409.231, "dur": 725.462,
    "args": {
      "Python parent id": 4674, "Python id": 4676, "finished": true, "Ev Idx": 11506
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857412.537, "dur": 0.344,
    "args": {
      "Python parent id": 4676, "Python id": 4677, "finished": true, "Ev Idx": 11507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_203", "pid": 537, "tid": 537,
    "ts": 7429564857415.052, "dur": 77.299,
    "args": {
      "Python parent id": 4676, "Python module id": 203, "Python id": 4678, "finished": true, "Ev Idx": 11508
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857415.539, "dur": 76.540,
    "args": {
      "Python parent id": 4678, "Python id": 4679, "finished": true, "Ev Idx": 11509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857415.691, "dur": 0.158,
    "args": {
      "Python parent id": 4679, "Python id": 4680, "finished": true, "Ev Idx": 11510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564857416.501, "dur": 75.400,
    "args": {
      "Python parent id": 4679, "Python id": 4681, "finished": true, "Ev Idx": 11511
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857416.811, "dur": 0.126,
    "args": {
      "Python parent id": 4681, "Python id": 4682, "finished": true, "Ev Idx": 11512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857417.119, "dur": 0.125,
    "args": {
      "Python parent id": 4681, "Python id": 4683, "finished": true, "Ev Idx": 11513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564857417.364, "dur": 74.423,
    "args": {
      "Python parent id": 4681, "Python id": 4684, "finished": true, "Ev Idx": 11514
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564857493.145, "dur": 3.869,
    "args": {
      "Python parent id": 4676, "Python id": 4685, "finished": true, "Ev Idx": 11515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564857497.423, "dur": 4.339,
    "args": {
      "Python parent id": 4676, "Python id": 4686, "finished": true, "Ev Idx": 11516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857502.740, "dur": 0.439,
    "args": {
      "Python parent id": 4676, "Python id": 4687, "finished": true, "Ev Idx": 11517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_204", "pid": 537, "tid": 537,
    "ts": 7429564857505.289, "dur": 48.206,
    "args": {
      "Python parent id": 4676, "Python module id": 204, "Python id": 4688, "finished": true, "Ev Idx": 11518
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857505.891, "dur": 47.369,
    "args": {
      "Python parent id": 4688, "Python id": 4689, "finished": true, "Ev Idx": 11519
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857506.180, "dur": 0.402,
    "args": {
      "Python parent id": 4689, "Python id": 4690, "finished": true, "Ev Idx": 11520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564857507.161, "dur": 45.990,
    "args": {
      "Python parent id": 4689, "Python id": 4691, "finished": true, "Ev Idx": 11521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857507.490, "dur": 0.124,
    "args": {
      "Python parent id": 4691, "Python id": 4692, "finished": true, "Ev Idx": 11522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857507.828, "dur": 0.118,
    "args": {
      "Python parent id": 4691, "Python id": 4693, "finished": true, "Ev Idx": 11523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564857508.228, "dur": 44.868,
    "args": {
      "Python parent id": 4691, "Python id": 4694, "finished": true, "Ev Idx": 11524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564857553.942, "dur": 2.526,
    "args": {
      "Python parent id": 4676, "Python id": 4695, "finished": true, "Ev Idx": 11525
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564857556.847, "dur": 2.550,
    "args": {
      "Python parent id": 4676, "Python id": 4696, "finished": true, "Ev Idx": 11526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857560.122, "dur": 0.460,
    "args": {
      "Python parent id": 4676, "Python id": 4697, "finished": true, "Ev Idx": 11527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_205", "pid": 537, "tid": 537,
    "ts": 7429564857562.625, "dur": 42.644,
    "args": {
      "Python parent id": 4676, "Python module id": 205, "Python id": 4698, "finished": true, "Ev Idx": 11528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564857563.100, "dur": 41.940,
    "args": {
      "Python parent id": 4698, "Python id": 4699, "finished": true, "Ev Idx": 11529
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564857563.292, "dur": 0.288,
    "args": {
      "Python parent id": 4699, "Python id": 4700, "finished": true, "Ev Idx": 11530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564857564.196, "dur": 40.698,
    "args": {
      "Python parent id": 4699, "Python id": 4701, "finished": true, "Ev Idx": 11531
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857564.474, "dur": 0.161,
    "args": {
      "Python parent id": 4701, "Python id": 4702, "finished": true, "Ev Idx": 11532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857564.834, "dur": 0.148,
    "args": {
      "Python parent id": 4701, "Python id": 4703, "finished": true, "Ev Idx": 11533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564857565.126, "dur": 39.715,
    "args": {
      "Python parent id": 4701, "Python id": 4704, "finished": true, "Ev Idx": 11534
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564857605.838, "dur": 2.658,
    "args": {
      "Python parent id": 4676, "Python id": 4705, "finished": true, "Ev Idx": 11535
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564857608.972, "dur": 2.502,
    "args": {
      "Python parent id": 4676, "Python id": 4706, "finished": true, "Ev Idx": 11536
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564857612.656, "dur": 150.251,
    "args": {
      "Python parent id": 4676, "Python id": 4707, "finished": true, "Ev Idx": 11537
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564857612.955, "dur": 3.919,
    "args": {
      "Python parent id": 4707, "Python id": 4708, "finished": true, "Ev Idx": 11538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564857617.183, "dur": 1.826,
    "args": {
      "Python parent id": 4707, "Python id": 4709, "finished": true, "Ev Idx": 11539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564857631.448, "dur": 51.481,
    "args": {
      "Python parent id": 4707, "Python id": 4710, "finished": true, "Ev Idx": 11540
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564857659.489, "dur": 22.349,
    "args": {
      "Python parent id": 4710, "Python id": 4711, "finished": true, "Ev Idx": 11541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564857713.983, "dur": 30.724,
    "args": {
      "Python parent id": 4707, "Python id": 4712, "finished": true, "Ev Idx": 11542
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564857730.654, "dur": 13.098,
    "args": {
      "Python parent id": 4712, "Python id": 4713, "finished": true, "Ev Idx": 11543
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564857767.164, "dur": 3.840,
    "args": {
      "Python parent id": 4676, "Python id": 4714, "finished": true, "Ev Idx": 11544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564857767.864, "dur": 0.389,
    "args": {
      "Python parent id": 4714, "Python id": 4715, "finished": true, "Ev Idx": 11545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564857768.640, "dur": 0.105,
    "args": {
      "Python parent id": 4714, "Python id": 4716, "finished": true, "Ev Idx": 11546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564857769.969, "dur": 0.269,
    "args": {
      "Python parent id": 4714, "Python id": 4717, "finished": true, "Ev Idx": 11547
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564857770.648, "dur": 0.113,
    "args": {
      "Python parent id": 4714, "Python id": 4718, "finished": true, "Ev Idx": 11548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564857772.327, "dur": 6.657,
    "args": {
      "Python parent id": 4676, "Python id": 4719, "finished": true, "Ev Idx": 11549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564857775.330, "dur": 3.490,
    "args": {
      "Python parent id": 4719, "Python id": 4720, "finished": true, "Ev Idx": 11550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564857775.594, "dur": 1.553,
    "args": {
      "Python parent id": 4720, "Python id": 4721, "finished": true, "Ev Idx": 11551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564857776.214, "dur": 0.787,
    "args": {
      "Python parent id": 4721, "Python id": 4722, "finished": true, "Ev Idx": 11552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564857777.431, "dur": 0.549,
    "args": {
      "Python parent id": 4720, "Python id": 4723, "finished": true, "Ev Idx": 11553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564857778.319, "dur": 0.429,
    "args": {
      "Python parent id": 4720, "Python id": 4724, "finished": true, "Ev Idx": 11554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564857781.708, "dur": 262.773,
    "args": {
      "Python parent id": 4676, "Python id": 4725, "finished": true, "Ev Idx": 11555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564857782.199, "dur": 48.083,
    "args": {
      "Python parent id": 4725, "Python id": 4726, "finished": true, "Ev Idx": 11556
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564857800.422, "dur": 5.024,
    "args": {
      "Python parent id": 4726, "Python id": 4727, "finished": true, "Ev Idx": 11557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564857806.305, "dur": 23.825,
    "args": {
      "Python parent id": 4726, "Python id": 4728, "finished": true, "Ev Idx": 11558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564857831.018, "dur": 32.598,
    "args": {
      "Python parent id": 4725, "Python id": 4729, "finished": true, "Ev Idx": 11559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564857848.544, "dur": 2.808,
    "args": {
      "Python parent id": 4729, "Python id": 4730, "finished": true, "Ev Idx": 11560
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564857851.915, "dur": 11.635,
    "args": {
      "Python parent id": 4729, "Python id": 4731, "finished": true, "Ev Idx": 11561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564857864.540, "dur": 3.309,
    "args": {
      "Python parent id": 4725, "Python id": 4732, "finished": true, "Ev Idx": 11562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564857868.356, "dur": 46.477,
    "args": {
      "Python parent id": 4725, "Python id": 4733, "finished": true, "Ev Idx": 11563
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564857947.903, "dur": 17.778,
    "args": {
      "Python parent id": 4725, "Python id": 4734, "finished": true, "Ev Idx": 11564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564857948.061, "dur": 0.301,
    "args": {
      "Python parent id": 4734, "Python id": 4735, "finished": true, "Ev Idx": 11565
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564857949.077, "dur": 16.421,
    "args": {
      "Python parent id": 4734, "Python id": 4736, "finished": true, "Ev Idx": 11566
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564857966.584, "dur": 1.757,
    "args": {
      "Python parent id": 4725, "Python id": 4737, "finished": true, "Ev Idx": 11567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564857970.018, "dur": 9.545,
    "args": {
      "Python parent id": 4725, "Python id": 4738, "finished": true, "Ev Idx": 11568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564857970.156, "dur": 0.116,
    "args": {
      "Python parent id": 4738, "Python id": 4739, "finished": true, "Ev Idx": 11569
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564857975.221, "dur": 0.982,
    "args": {
      "Python parent id": 4738, "Python id": 4740, "finished": true, "Ev Idx": 11570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564857975.683, "dur": 0.438,
    "args": {
      "Python parent id": 4740, "Python id": 4741, "finished": true, "Ev Idx": 11571
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564857976.615, "dur": 2.865,
    "args": {
      "Python parent id": 4738, "Python id": 4742, "finished": true, "Ev Idx": 11572
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564857980.200, "dur": 44.207,
    "args": {
      "Python parent id": 4725, "Python id": 4743, "finished": true, "Ev Idx": 11573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564858025.021, "dur": 3.008,
    "args": {
      "Python parent id": 4725, "Python id": 4744, "finished": true, "Ev Idx": 11574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564858028.779, "dur": 14.610,
    "args": {
      "Python parent id": 4725, "Python id": 4745, "finished": true, "Ev Idx": 11575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564858047.462, "dur": 4.186,
    "args": {
      "Python parent id": 4676, "Python id": 4746, "finished": true, "Ev Idx": 11576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564858052.377, "dur": 0.264,
    "args": {
      "Python parent id": 4676, "Python id": 4747, "finished": true, "Ev Idx": 11577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858053.612, "dur": 0.932,
    "args": {
      "Python parent id": 4676, "Python id": 4748, "finished": true, "Ev Idx": 11578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_206", "pid": 537, "tid": 537,
    "ts": 7429564858058.931, "dur": 74.816,
    "args": {
      "Python parent id": 4676, "Python module id": 206, "Python id": 4749, "finished": true, "Ev Idx": 11579
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858060.052, "dur": 73.149,
    "args": {
      "Python parent id": 4749, "Python id": 4750, "finished": true, "Ev Idx": 11580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858060.505, "dur": 1.091,
    "args": {
      "Python parent id": 4750, "Python id": 4751, "finished": true, "Ev Idx": 11581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564858062.710, "dur": 70.140,
    "args": {
      "Python parent id": 4750, "Python id": 4752, "finished": true, "Ev Idx": 11582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858063.118, "dur": 0.282,
    "args": {
      "Python parent id": 4752, "Python id": 4753, "finished": true, "Ev Idx": 11583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858063.708, "dur": 0.165,
    "args": {
      "Python parent id": 4752, "Python id": 4754, "finished": true, "Ev Idx": 11584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564858064.034, "dur": 68.717,
    "args": {
      "Python parent id": 4752, "Python id": 4755, "finished": true, "Ev Idx": 11585
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858150.506, "dur": 0.803,
    "args": {
      "Python parent id": 4660, "Python id": 4756, "finished": true, "Ev Idx": 11586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_59", "pid": 537, "tid": 537,
    "ts": 7429564858154.591, "dur": 91.876,
    "args": {
      "Python parent id": 4660, "Python module id": 59, "Python id": 4757, "finished": true, "Ev Idx": 11587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858155.425, "dur": 90.686,
    "args": {
      "Python parent id": 4757, "Python id": 4758, "finished": true, "Ev Idx": 11588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858155.711, "dur": 0.548,
    "args": {
      "Python parent id": 4758, "Python id": 4759, "finished": true, "Ev Idx": 11589
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564858157.557, "dur": 87.545,
    "args": {
      "Python parent id": 4758, "Python id": 4760, "finished": true, "Ev Idx": 11590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564858158.292, "dur": 2.007,
    "args": {
      "Python parent id": 4760, "Python id": 4761, "finished": true, "Ev Idx": 11591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564858160.642, "dur": 15.616,
    "args": {
      "Python parent id": 4760, "Python id": 4762, "finished": true, "Ev Idx": 11592
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564858176.682, "dur": 17.612,
    "args": {
      "Python parent id": 4760, "Python id": 4763, "finished": true, "Ev Idx": 11593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564858208.869, "dur": 11.065,
    "args": {
      "Python parent id": 4760, "Python id": 4764, "finished": true, "Ev Idx": 11594
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858232.406, "dur": 0.191,
    "args": {
      "Python parent id": 4760, "Python id": 4765, "finished": true, "Ev Idx": 11595
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564858233.009, "dur": 1.569,
    "args": {
      "Python parent id": 4760, "Python id": 4766, "finished": true, "Ev Idx": 11596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858247.248, "dur": 0.394,
    "args": {
      "Python parent id": 4660, "Python id": 4767, "finished": true, "Ev Idx": 11597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_29", "pid": 537, "tid": 537,
    "ts": 7429564858249.418, "dur": 296.813,
    "args": {
      "Python parent id": 4660, "Python module id": 29, "Python id": 4768, "finished": true, "Ev Idx": 11598
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858249.961, "dur": 296.063,
    "args": {
      "Python parent id": 4768, "Python id": 4769, "finished": true, "Ev Idx": 11599
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858250.166, "dur": 0.403,
    "args": {
      "Python parent id": 4769, "Python id": 4770, "finished": true, "Ev Idx": 11600
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564858251.712, "dur": 294.173,
    "args": {
      "Python parent id": 4769, "Python id": 4771, "finished": true, "Ev Idx": 11601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858252.035, "dur": 0.324,
    "args": {
      "Python parent id": 4771, "Python id": 4772, "finished": true, "Ev Idx": 11602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858252.688, "dur": 0.310,
    "args": {
      "Python parent id": 4771, "Python id": 4773, "finished": true, "Ev Idx": 11603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858253.227, "dur": 0.246,
    "args": {
      "Python parent id": 4771, "Python id": 4774, "finished": true, "Ev Idx": 11604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_207", "pid": 537, "tid": 537,
    "ts": 7429564858255.694, "dur": 122.721,
    "args": {
      "Python parent id": 4771, "Python module id": 207, "Python id": 4775, "finished": true, "Ev Idx": 11605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858256.247, "dur": 121.784,
    "args": {
      "Python parent id": 4775, "Python id": 4776, "finished": true, "Ev Idx": 11606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858256.384, "dur": 0.174,
    "args": {
      "Python parent id": 4776, "Python id": 4777, "finished": true, "Ev Idx": 11607
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564858257.227, "dur": 120.588,
    "args": {
      "Python parent id": 4776, "Python id": 4778, "finished": true, "Ev Idx": 11608
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858257.512, "dur": 0.150,
    "args": {
      "Python parent id": 4778, "Python id": 4779, "finished": true, "Ev Idx": 11609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858257.915, "dur": 0.169,
    "args": {
      "Python parent id": 4778, "Python id": 4780, "finished": true, "Ev Idx": 11610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564858258.239, "dur": 119.460,
    "args": {
      "Python parent id": 4778, "Python id": 4781, "finished": true, "Ev Idx": 11611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_29", "pid": 537, "tid": 537,
    "ts": 7429564858380.622, "dur": 20.702,
    "args": {
      "Python parent id": 4771, "Python module id": 29, "Python id": 4782, "finished": true, "Ev Idx": 11612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858381.316, "dur": 19.794,
    "args": {
      "Python parent id": 4782, "Python id": 4783, "finished": true, "Ev Idx": 11613
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858381.605, "dur": 0.493,
    "args": {
      "Python parent id": 4783, "Python id": 4784, "finished": true, "Ev Idx": 11614
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564858383.341, "dur": 17.671,
    "args": {
      "Python parent id": 4783, "Python id": 4785, "finished": true, "Ev Idx": 11615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564858384.220, "dur": 16.639,
    "args": {
      "Python parent id": 4785, "Python id": 4786, "finished": true, "Ev Idx": 11616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564858384.364, "dur": 0.294,
    "args": {
      "Python parent id": 4786, "Python id": 4787, "finished": true, "Ev Idx": 11617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564858385.594, "dur": 15.140,
    "args": {
      "Python parent id": 4786, "Python id": 4788, "finished": true, "Ev Idx": 11618
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858402.268, "dur": 0.581,
    "args": {
      "Python parent id": 4771, "Python id": 4789, "finished": true, "Ev Idx": 11619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_208", "pid": 537, "tid": 537,
    "ts": 7429564858405.306, "dur": 80.949,
    "args": {
      "Python parent id": 4771, "Python module id": 208, "Python id": 4790, "finished": true, "Ev Idx": 11620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858405.910, "dur": 80.056,
    "args": {
      "Python parent id": 4790, "Python id": 4791, "finished": true, "Ev Idx": 11621
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858406.121, "dur": 0.266,
    "args": {
      "Python parent id": 4791, "Python id": 4792, "finished": true, "Ev Idx": 11622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564858407.234, "dur": 78.520,
    "args": {
      "Python parent id": 4791, "Python id": 4793, "finished": true, "Ev Idx": 11623
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858407.571, "dur": 0.174,
    "args": {
      "Python parent id": 4793, "Python id": 4794, "finished": true, "Ev Idx": 11624
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858407.927, "dur": 0.143,
    "args": {
      "Python parent id": 4793, "Python id": 4795, "finished": true, "Ev Idx": 11625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564858408.197, "dur": 77.467,
    "args": {
      "Python parent id": 4793, "Python id": 4796, "finished": true, "Ev Idx": 11626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_209", "pid": 537, "tid": 537,
    "ts": 7429564858500.065, "dur": 45.498,
    "args": {
      "Python parent id": 4771, "Python module id": 209, "Python id": 4797, "finished": true, "Ev Idx": 11627
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858500.794, "dur": 44.467,
    "args": {
      "Python parent id": 4797, "Python id": 4798, "finished": true, "Ev Idx": 11628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858501.143, "dur": 0.444,
    "args": {
      "Python parent id": 4798, "Python id": 4799, "finished": true, "Ev Idx": 11629
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564858502.269, "dur": 42.877,
    "args": {
      "Python parent id": 4798, "Python id": 4800, "finished": true, "Ev Idx": 11630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858502.588, "dur": 0.259,
    "args": {
      "Python parent id": 4800, "Python id": 4801, "finished": true, "Ev Idx": 11631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858503.048, "dur": 0.121,
    "args": {
      "Python parent id": 4800, "Python id": 4802, "finished": true, "Ev Idx": 11632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564858503.299, "dur": 41.756,
    "args": {
      "Python parent id": 4800, "Python id": 4803, "finished": true, "Ev Idx": 11633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_30", "pid": 537, "tid": 537,
    "ts": 7429564858567.301, "dur": 1183.748,
    "args": {
      "Python parent id": 64, "Python module id": 30, "Python id": 4804, "finished": true, "Ev Idx": 11634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858569.039, "dur": 1180.963,
    "args": {
      "Python parent id": 4804, "Python id": 4805, "finished": true, "Ev Idx": 11635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858569.311, "dur": 0.381,
    "args": {
      "Python parent id": 4805, "Python id": 4806, "finished": true, "Ev Idx": 11636
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564858571.419, "dur": 1177.645,
    "args": {
      "Python parent id": 4805, "Python id": 4807, "finished": true, "Ev Idx": 11637
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858571.928, "dur": 0.365,
    "args": {
      "Python parent id": 4807, "Python id": 4808, "finished": true, "Ev Idx": 11638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_60", "pid": 537, "tid": 537,
    "ts": 7429564858574.299, "dur": 79.576,
    "args": {
      "Python parent id": 4807, "Python module id": 60, "Python id": 4809, "finished": true, "Ev Idx": 11639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858574.773, "dur": 78.810,
    "args": {
      "Python parent id": 4809, "Python id": 4810, "finished": true, "Ev Idx": 11640
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858574.919, "dur": 0.154,
    "args": {
      "Python parent id": 4810, "Python id": 4811, "finished": true, "Ev Idx": 11641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564858575.962, "dur": 76.816,
    "args": {
      "Python parent id": 4810, "Python id": 4812, "finished": true, "Ev Idx": 11642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564858576.602, "dur": 2.137,
    "args": {
      "Python parent id": 4812, "Python id": 4813, "finished": true, "Ev Idx": 11643
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564858579.065, "dur": 12.289,
    "args": {
      "Python parent id": 4812, "Python id": 4814, "finished": true, "Ev Idx": 11644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564858591.691, "dur": 14.895,
    "args": {
      "Python parent id": 4812, "Python id": 4815, "finished": true, "Ev Idx": 11645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564858619.714, "dur": 10.350,
    "args": {
      "Python parent id": 4812, "Python id": 4816, "finished": true, "Ev Idx": 11646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858641.037, "dur": 0.234,
    "args": {
      "Python parent id": 4812, "Python id": 4817, "finished": true, "Ev Idx": 11647
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564858641.745, "dur": 1.398,
    "args": {
      "Python parent id": 4812, "Python id": 4818, "finished": true, "Ev Idx": 11648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858654.460, "dur": 0.390,
    "args": {
      "Python parent id": 4807, "Python id": 4819, "finished": true, "Ev Idx": 11649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_30", "pid": 537, "tid": 537,
    "ts": 7429564858658.021, "dur": 697.758,
    "args": {
      "Python parent id": 4807, "Python module id": 30, "Python id": 4820, "finished": true, "Ev Idx": 11650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858659.619, "dur": 695.511,
    "args": {
      "Python parent id": 4820, "Python id": 4821, "finished": true, "Ev Idx": 11651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858659.787, "dur": 0.299,
    "args": {
      "Python parent id": 4821, "Python id": 4822, "finished": true, "Ev Idx": 11652
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564858661.746, "dur": 692.697,
    "args": {
      "Python parent id": 4821, "Python id": 4823, "finished": true, "Ev Idx": 11653
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858665.314, "dur": 0.297,
    "args": {
      "Python parent id": 4823, "Python id": 4824, "finished": true, "Ev Idx": 11654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_210", "pid": 537, "tid": 537,
    "ts": 7429564858667.503, "dur": 40.194,
    "args": {
      "Python parent id": 4823, "Python module id": 210, "Python id": 4825, "finished": true, "Ev Idx": 11655
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858668.114, "dur": 39.270,
    "args": {
      "Python parent id": 4825, "Python id": 4826, "finished": true, "Ev Idx": 11656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858668.288, "dur": 0.168,
    "args": {
      "Python parent id": 4826, "Python id": 4827, "finished": true, "Ev Idx": 11657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564858669.120, "dur": 38.091,
    "args": {
      "Python parent id": 4826, "Python id": 4828, "finished": true, "Ev Idx": 11658
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858669.474, "dur": 0.146,
    "args": {
      "Python parent id": 4828, "Python id": 4829, "finished": true, "Ev Idx": 11659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858669.856, "dur": 0.172,
    "args": {
      "Python parent id": 4828, "Python id": 4830, "finished": true, "Ev Idx": 11660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564858670.130, "dur": 36.974,
    "args": {
      "Python parent id": 4828, "Python id": 4831, "finished": true, "Ev Idx": 11661
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564858708.383, "dur": 3.715,
    "args": {
      "Python parent id": 4823, "Python id": 4832, "finished": true, "Ev Idx": 11662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564858712.550, "dur": 3.855,
    "args": {
      "Python parent id": 4823, "Python id": 4833, "finished": true, "Ev Idx": 11663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858717.228, "dur": 0.352,
    "args": {
      "Python parent id": 4823, "Python id": 4834, "finished": true, "Ev Idx": 11664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_211", "pid": 537, "tid": 537,
    "ts": 7429564858719.328, "dur": 43.596,
    "args": {
      "Python parent id": 4823, "Python module id": 211, "Python id": 4835, "finished": true, "Ev Idx": 11665
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858719.830, "dur": 42.816,
    "args": {
      "Python parent id": 4835, "Python id": 4836, "finished": true, "Ev Idx": 11666
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858720.092, "dur": 0.255,
    "args": {
      "Python parent id": 4836, "Python id": 4837, "finished": true, "Ev Idx": 11667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564858721.050, "dur": 41.492,
    "args": {
      "Python parent id": 4836, "Python id": 4838, "finished": true, "Ev Idx": 11668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858721.331, "dur": 0.138,
    "args": {
      "Python parent id": 4838, "Python id": 4839, "finished": true, "Ev Idx": 11669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858721.637, "dur": 0.140,
    "args": {
      "Python parent id": 4838, "Python id": 4840, "finished": true, "Ev Idx": 11670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564858721.883, "dur": 40.565,
    "args": {
      "Python parent id": 4838, "Python id": 4841, "finished": true, "Ev Idx": 11671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564858763.391, "dur": 2.490,
    "args": {
      "Python parent id": 4823, "Python id": 4842, "finished": true, "Ev Idx": 11672
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564858766.420, "dur": 3.439,
    "args": {
      "Python parent id": 4823, "Python id": 4843, "finished": true, "Ev Idx": 11673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858770.580, "dur": 0.439,
    "args": {
      "Python parent id": 4823, "Python id": 4844, "finished": true, "Ev Idx": 11674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_212", "pid": 537, "tid": 537,
    "ts": 7429564858773.148, "dur": 40.625,
    "args": {
      "Python parent id": 4823, "Python module id": 212, "Python id": 4845, "finished": true, "Ev Idx": 11675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564858773.788, "dur": 39.680,
    "args": {
      "Python parent id": 4845, "Python id": 4846, "finished": true, "Ev Idx": 11676
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564858774.047, "dur": 0.293,
    "args": {
      "Python parent id": 4846, "Python id": 4847, "finished": true, "Ev Idx": 11677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564858774.992, "dur": 38.349,
    "args": {
      "Python parent id": 4846, "Python id": 4848, "finished": true, "Ev Idx": 11678
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858775.244, "dur": 0.169,
    "args": {
      "Python parent id": 4848, "Python id": 4849, "finished": true, "Ev Idx": 11679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564858775.618, "dur": 0.123,
    "args": {
      "Python parent id": 4848, "Python id": 4850, "finished": true, "Ev Idx": 11680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564858775.841, "dur": 37.428,
    "args": {
      "Python parent id": 4848, "Python id": 4851, "finished": true, "Ev Idx": 11681
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564858814.375, "dur": 2.443,
    "args": {
      "Python parent id": 4823, "Python id": 4852, "finished": true, "Ev Idx": 11682
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564858817.358, "dur": 2.368,
    "args": {
      "Python parent id": 4823, "Python id": 4853, "finished": true, "Ev Idx": 11683
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564858820.891, "dur": 196.233,
    "args": {
      "Python parent id": 4823, "Python id": 4854, "finished": true, "Ev Idx": 11684
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564858821.221, "dur": 3.768,
    "args": {
      "Python parent id": 4854, "Python id": 4855, "finished": true, "Ev Idx": 11685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564858825.321, "dur": 1.796,
    "args": {
      "Python parent id": 4854, "Python id": 4856, "finished": true, "Ev Idx": 11686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564858838.797, "dur": 51.893,
    "args": {
      "Python parent id": 4854, "Python id": 4857, "finished": true, "Ev Idx": 11687
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564858866.415, "dur": 23.150,
    "args": {
      "Python parent id": 4857, "Python id": 4858, "finished": true, "Ev Idx": 11688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564858965.123, "dur": 33.404,
    "args": {
      "Python parent id": 4854, "Python id": 4859, "finished": true, "Ev Idx": 11689
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564858985.110, "dur": 12.588,
    "args": {
      "Python parent id": 4859, "Python id": 4860, "finished": true, "Ev Idx": 11690
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564859021.801, "dur": 4.104,
    "args": {
      "Python parent id": 4823, "Python id": 4861, "finished": true, "Ev Idx": 11691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564859022.547, "dur": 0.370,
    "args": {
      "Python parent id": 4861, "Python id": 4862, "finished": true, "Ev Idx": 11692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564859023.339, "dur": 0.049,
    "args": {
      "Python parent id": 4861, "Python id": 4863, "finished": true, "Ev Idx": 11693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564859024.801, "dur": 0.279,
    "args": {
      "Python parent id": 4861, "Python id": 4864, "finished": true, "Ev Idx": 11694
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564859025.490, "dur": 0.118,
    "args": {
      "Python parent id": 4861, "Python id": 4865, "finished": true, "Ev Idx": 11695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564859027.420, "dur": 6.789,
    "args": {
      "Python parent id": 4823, "Python id": 4866, "finished": true, "Ev Idx": 11696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564859030.723, "dur": 3.323,
    "args": {
      "Python parent id": 4866, "Python id": 4867, "finished": true, "Ev Idx": 11697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564859030.940, "dur": 1.569,
    "args": {
      "Python parent id": 4867, "Python id": 4868, "finished": true, "Ev Idx": 11698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564859031.412, "dur": 0.913,
    "args": {
      "Python parent id": 4868, "Python id": 4869, "finished": true, "Ev Idx": 11699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564859032.810, "dur": 0.501,
    "args": {
      "Python parent id": 4867, "Python id": 4870, "finished": true, "Ev Idx": 11700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564859033.546, "dur": 0.428,
    "args": {
      "Python parent id": 4867, "Python id": 4871, "finished": true, "Ev Idx": 11701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564859037.041, "dur": 259.913,
    "args": {
      "Python parent id": 4823, "Python id": 4872, "finished": true, "Ev Idx": 11702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564859037.646, "dur": 48.137,
    "args": {
      "Python parent id": 4872, "Python id": 4873, "finished": true, "Ev Idx": 11703
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564859056.106, "dur": 5.419,
    "args": {
      "Python parent id": 4873, "Python id": 4874, "finished": true, "Ev Idx": 11704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564859062.457, "dur": 23.170,
    "args": {
      "Python parent id": 4873, "Python id": 4875, "finished": true, "Ev Idx": 11705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564859086.663, "dur": 32.206,
    "args": {
      "Python parent id": 4872, "Python id": 4876, "finished": true, "Ev Idx": 11706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564859100.846, "dur": 2.841,
    "args": {
      "Python parent id": 4876, "Python id": 4877, "finished": true, "Ev Idx": 11707
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564859104.233, "dur": 14.571,
    "args": {
      "Python parent id": 4876, "Python id": 4878, "finished": true, "Ev Idx": 11708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564859119.844, "dur": 3.099,
    "args": {
      "Python parent id": 4872, "Python id": 4879, "finished": true, "Ev Idx": 11709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564859123.364, "dur": 48.127,
    "args": {
      "Python parent id": 4872, "Python id": 4880, "finished": true, "Ev Idx": 11710
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564859204.389, "dur": 17.321,
    "args": {
      "Python parent id": 4872, "Python id": 4881, "finished": true, "Ev Idx": 11711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564859204.537, "dur": 0.230,
    "args": {
      "Python parent id": 4881, "Python id": 4882, "finished": true, "Ev Idx": 11712
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564859205.538, "dur": 16.023,
    "args": {
      "Python parent id": 4881, "Python id": 4883, "finished": true, "Ev Idx": 11713
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564859222.645, "dur": 1.701,
    "args": {
      "Python parent id": 4872, "Python id": 4884, "finished": true, "Ev Idx": 11714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564859225.931, "dur": 9.705,
    "args": {
      "Python parent id": 4872, "Python id": 4885, "finished": true, "Ev Idx": 11715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564859226.106, "dur": 0.054,
    "args": {
      "Python parent id": 4885, "Python id": 4886, "finished": true, "Ev Idx": 11716
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859231.460, "dur": 1.045,
    "args": {
      "Python parent id": 4885, "Python id": 4887, "finished": true, "Ev Idx": 11717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564859231.994, "dur": 0.430,
    "args": {
      "Python parent id": 4887, "Python id": 4888, "finished": true, "Ev Idx": 11718
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564859232.891, "dur": 2.657,
    "args": {
      "Python parent id": 4885, "Python id": 4889, "finished": true, "Ev Idx": 11719
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564859236.258, "dur": 42.255,
    "args": {
      "Python parent id": 4872, "Python id": 4890, "finished": true, "Ev Idx": 11720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564859279.110, "dur": 2.978,
    "args": {
      "Python parent id": 4872, "Python id": 4891, "finished": true, "Ev Idx": 11721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564859282.751, "dur": 13.204,
    "args": {
      "Python parent id": 4872, "Python id": 4892, "finished": true, "Ev Idx": 11722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564859299.661, "dur": 3.380,
    "args": {
      "Python parent id": 4823, "Python id": 4893, "finished": true, "Ev Idx": 11723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564859303.689, "dur": 0.247,
    "args": {
      "Python parent id": 4823, "Python id": 4894, "finished": true, "Ev Idx": 11724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859304.896, "dur": 0.988,
    "args": {
      "Python parent id": 4823, "Python id": 4895, "finished": true, "Ev Idx": 11725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_213", "pid": 537, "tid": 537,
    "ts": 7429564859310.130, "dur": 43.677,
    "args": {
      "Python parent id": 4823, "Python module id": 213, "Python id": 4896, "finished": true, "Ev Idx": 11726
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859311.132, "dur": 42.301,
    "args": {
      "Python parent id": 4896, "Python id": 4897, "finished": true, "Ev Idx": 11727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859311.680, "dur": 1.200,
    "args": {
      "Python parent id": 4897, "Python id": 4898, "finished": true, "Ev Idx": 11728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564859314.052, "dur": 39.129,
    "args": {
      "Python parent id": 4897, "Python id": 4899, "finished": true, "Ev Idx": 11729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859314.385, "dur": 0.228,
    "args": {
      "Python parent id": 4899, "Python id": 4900, "finished": true, "Ev Idx": 11730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859314.863, "dur": 0.174,
    "args": {
      "Python parent id": 4899, "Python id": 4901, "finished": true, "Ev Idx": 11731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564859315.181, "dur": 37.897,
    "args": {
      "Python parent id": 4899, "Python id": 4902, "finished": true, "Ev Idx": 11732
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859368.831, "dur": 0.620,
    "args": {
      "Python parent id": 4807, "Python id": 4903, "finished": true, "Ev Idx": 11733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_61", "pid": 537, "tid": 537,
    "ts": 7429564859372.089, "dur": 83.860,
    "args": {
      "Python parent id": 4807, "Python module id": 61, "Python id": 4904, "finished": true, "Ev Idx": 11734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859372.749, "dur": 82.866,
    "args": {
      "Python parent id": 4904, "Python id": 4905, "finished": true, "Ev Idx": 11735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859372.957, "dur": 0.493,
    "args": {
      "Python parent id": 4905, "Python id": 4906, "finished": true, "Ev Idx": 11736
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564859374.296, "dur": 80.323,
    "args": {
      "Python parent id": 4905, "Python id": 4907, "finished": true, "Ev Idx": 11737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564859374.857, "dur": 1.937,
    "args": {
      "Python parent id": 4907, "Python id": 4908, "finished": true, "Ev Idx": 11738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564859377.080, "dur": 13.847,
    "args": {
      "Python parent id": 4907, "Python id": 4909, "finished": true, "Ev Idx": 11739
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564859391.298, "dur": 17.040,
    "args": {
      "Python parent id": 4907, "Python id": 4910, "finished": true, "Ev Idx": 11740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564859419.830, "dur": 13.410,
    "args": {
      "Python parent id": 4907, "Python id": 4911, "finished": true, "Ev Idx": 11741
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859444.259, "dur": 0.203,
    "args": {
      "Python parent id": 4907, "Python id": 4912, "finished": true, "Ev Idx": 11742
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564859444.819, "dur": 1.172,
    "args": {
      "Python parent id": 4907, "Python id": 4913, "finished": true, "Ev Idx": 11743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859456.676, "dur": 0.585,
    "args": {
      "Python parent id": 4807, "Python id": 4914, "finished": true, "Ev Idx": 11744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_30", "pid": 537, "tid": 537,
    "ts": 7429564859458.865, "dur": 278.666,
    "args": {
      "Python parent id": 4807, "Python module id": 30, "Python id": 4915, "finished": true, "Ev Idx": 11745
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859459.346, "dur": 277.909,
    "args": {
      "Python parent id": 4915, "Python id": 4916, "finished": true, "Ev Idx": 11746
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859459.551, "dur": 0.379,
    "args": {
      "Python parent id": 4916, "Python id": 4917, "finished": true, "Ev Idx": 11747
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564859461.014, "dur": 276.093,
    "args": {
      "Python parent id": 4916, "Python id": 4918, "finished": true, "Ev Idx": 11748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859461.274, "dur": 0.338,
    "args": {
      "Python parent id": 4918, "Python id": 4919, "finished": true, "Ev Idx": 11749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859461.991, "dur": 0.346,
    "args": {
      "Python parent id": 4918, "Python id": 4920, "finished": true, "Ev Idx": 11750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859462.577, "dur": 0.256,
    "args": {
      "Python parent id": 4918, "Python id": 4921, "finished": true, "Ev Idx": 11751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_214", "pid": 537, "tid": 537,
    "ts": 7429564859464.488, "dur": 49.036,
    "args": {
      "Python parent id": 4918, "Python module id": 214, "Python id": 4922, "finished": true, "Ev Idx": 11752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859465.243, "dur": 48.011,
    "args": {
      "Python parent id": 4922, "Python id": 4923, "finished": true, "Ev Idx": 11753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859465.406, "dur": 0.156,
    "args": {
      "Python parent id": 4923, "Python id": 4924, "finished": true, "Ev Idx": 11754
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564859466.288, "dur": 46.834,
    "args": {
      "Python parent id": 4923, "Python id": 4925, "finished": true, "Ev Idx": 11755
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859466.570, "dur": 0.148,
    "args": {
      "Python parent id": 4925, "Python id": 4926, "finished": true, "Ev Idx": 11756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859466.961, "dur": 0.141,
    "args": {
      "Python parent id": 4925, "Python id": 4927, "finished": true, "Ev Idx": 11757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564859467.209, "dur": 45.804,
    "args": {
      "Python parent id": 4925, "Python id": 4928, "finished": true, "Ev Idx": 11758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_30", "pid": 537, "tid": 537,
    "ts": 7429564859515.747, "dur": 68.212,
    "args": {
      "Python parent id": 4918, "Python module id": 30, "Python id": 4929, "finished": true, "Ev Idx": 11759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859516.247, "dur": 67.381,
    "args": {
      "Python parent id": 4929, "Python id": 4930, "finished": true, "Ev Idx": 11760
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859516.440, "dur": 0.253,
    "args": {
      "Python parent id": 4930, "Python id": 4931, "finished": true, "Ev Idx": 11761
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564859517.838, "dur": 65.625,
    "args": {
      "Python parent id": 4930, "Python id": 4932, "finished": true, "Ev Idx": 11762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564859518.699, "dur": 64.542,
    "args": {
      "Python parent id": 4932, "Python id": 4933, "finished": true, "Ev Idx": 11763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564859518.853, "dur": 0.185,
    "args": {
      "Python parent id": 4933, "Python id": 4934, "finished": true, "Ev Idx": 11764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564859519.984, "dur": 63.127,
    "args": {
      "Python parent id": 4933, "Python id": 4935, "finished": true, "Ev Idx": 11765
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859584.819, "dur": 0.410,
    "args": {
      "Python parent id": 4918, "Python id": 4936, "finished": true, "Ev Idx": 11766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_215", "pid": 537, "tid": 537,
    "ts": 7429564859587.258, "dur": 48.385,
    "args": {
      "Python parent id": 4918, "Python module id": 215, "Python id": 4937, "finished": true, "Ev Idx": 11767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859587.736, "dur": 47.658,
    "args": {
      "Python parent id": 4937, "Python id": 4938, "finished": true, "Ev Idx": 11768
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859587.993, "dur": 0.325,
    "args": {
      "Python parent id": 4938, "Python id": 4939, "finished": true, "Ev Idx": 11769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564859589.051, "dur": 46.218,
    "args": {
      "Python parent id": 4938, "Python id": 4940, "finished": true, "Ev Idx": 11770
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859589.368, "dur": 0.121,
    "args": {
      "Python parent id": 4940, "Python id": 4941, "finished": true, "Ev Idx": 11771
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859589.667, "dur": 0.316,
    "args": {
      "Python parent id": 4940, "Python id": 4942, "finished": true, "Ev Idx": 11772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564859590.087, "dur": 45.111,
    "args": {
      "Python parent id": 4940, "Python id": 4943, "finished": true, "Ev Idx": 11773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_216", "pid": 537, "tid": 537,
    "ts": 7429564859685.970, "dur": 50.761,
    "args": {
      "Python parent id": 4918, "Python module id": 216, "Python id": 4944, "finished": true, "Ev Idx": 11774
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859686.825, "dur": 49.585,
    "args": {
      "Python parent id": 4944, "Python id": 4945, "finished": true, "Ev Idx": 11775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859687.064, "dur": 0.477,
    "args": {
      "Python parent id": 4945, "Python id": 4946, "finished": true, "Ev Idx": 11776
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564859688.527, "dur": 47.784,
    "args": {
      "Python parent id": 4945, "Python id": 4947, "finished": true, "Ev Idx": 11777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859688.939, "dur": 0.233,
    "args": {
      "Python parent id": 4947, "Python id": 4948, "finished": true, "Ev Idx": 11778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859689.378, "dur": 0.143,
    "args": {
      "Python parent id": 4947, "Python id": 4949, "finished": true, "Ev Idx": 11779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564859689.635, "dur": 46.591,
    "args": {
      "Python parent id": 4947, "Python id": 4950, "finished": true, "Ev Idx": 11780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_31", "pid": 537, "tid": 537,
    "ts": 7429564859757.453, "dur": 1161.545,
    "args": {
      "Python parent id": 64, "Python module id": 31, "Python id": 4951, "finished": true, "Ev Idx": 11781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859758.979, "dur": 1159.181,
    "args": {
      "Python parent id": 4951, "Python id": 4952, "finished": true, "Ev Idx": 11782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859759.161, "dur": 0.382,
    "args": {
      "Python parent id": 4952, "Python id": 4953, "finished": true, "Ev Idx": 11783
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(301): forward", "pid": 537, "tid": 537,
    "ts": 7429564859761.333, "dur": 1155.989,
    "args": {
      "Python parent id": 4952, "Python id": 4954, "finished": true, "Ev Idx": 11784
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859761.860, "dur": 0.481,
    "args": {
      "Python parent id": 4954, "Python id": 4955, "finished": true, "Ev Idx": 11785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_62", "pid": 537, "tid": 537,
    "ts": 7429564859764.882, "dur": 77.759,
    "args": {
      "Python parent id": 4954, "Python module id": 62, "Python id": 4956, "finished": true, "Ev Idx": 11786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859765.477, "dur": 76.859,
    "args": {
      "Python parent id": 4956, "Python id": 4957, "finished": true, "Ev Idx": 11787
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859765.619, "dur": 0.358,
    "args": {
      "Python parent id": 4957, "Python id": 4958, "finished": true, "Ev Idx": 11788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564859766.680, "dur": 74.761,
    "args": {
      "Python parent id": 4957, "Python id": 4959, "finished": true, "Ev Idx": 11789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564859767.323, "dur": 2.170,
    "args": {
      "Python parent id": 4959, "Python id": 4960, "finished": true, "Ev Idx": 11790
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564859769.801, "dur": 12.022,
    "args": {
      "Python parent id": 4959, "Python id": 4961, "finished": true, "Ev Idx": 11791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564859782.197, "dur": 15.665,
    "args": {
      "Python parent id": 4959, "Python id": 4962, "finished": true, "Ev Idx": 11792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564859810.105, "dur": 9.414,
    "args": {
      "Python parent id": 4959, "Python id": 4963, "finished": true, "Ev Idx": 11793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859830.659, "dur": 0.191,
    "args": {
      "Python parent id": 4959, "Python id": 4964, "finished": true, "Ev Idx": 11794
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564859831.275, "dur": 1.187,
    "args": {
      "Python parent id": 4959, "Python id": 4965, "finished": true, "Ev Idx": 11795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859843.281, "dur": 0.584,
    "args": {
      "Python parent id": 4954, "Python id": 4966, "finished": true, "Ev Idx": 11796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_31", "pid": 537, "tid": 537,
    "ts": 7429564859846.681, "dur": 670.088,
    "args": {
      "Python parent id": 4954, "Python module id": 31, "Python id": 4967, "finished": true, "Ev Idx": 11797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859848.125, "dur": 668.010,
    "args": {
      "Python parent id": 4967, "Python id": 4968, "finished": true, "Ev Idx": 11798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859848.360, "dur": 0.350,
    "args": {
      "Python parent id": 4968, "Python id": 4969, "finished": true, "Ev Idx": 11799
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(240): forward", "pid": 537, "tid": 537,
    "ts": 7429564859850.379, "dur": 664.873,
    "args": {
      "Python parent id": 4968, "Python id": 4970, "finished": true, "Ev Idx": 11800
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859853.805, "dur": 0.319,
    "args": {
      "Python parent id": 4970, "Python id": 4971, "finished": true, "Ev Idx": 11801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_217", "pid": 537, "tid": 537,
    "ts": 7429564859856.395, "dur": 43.353,
    "args": {
      "Python parent id": 4970, "Python module id": 217, "Python id": 4972, "finished": true, "Ev Idx": 11802
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859856.951, "dur": 42.569,
    "args": {
      "Python parent id": 4972, "Python id": 4973, "finished": true, "Ev Idx": 11803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859857.119, "dur": 0.172,
    "args": {
      "Python parent id": 4973, "Python id": 4974, "finished": true, "Ev Idx": 11804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564859858.209, "dur": 41.143,
    "args": {
      "Python parent id": 4973, "Python id": 4975, "finished": true, "Ev Idx": 11805
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859858.553, "dur": 0.168,
    "args": {
      "Python parent id": 4975, "Python id": 4976, "finished": true, "Ev Idx": 11806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859858.996, "dur": 0.133,
    "args": {
      "Python parent id": 4975, "Python id": 4977, "finished": true, "Ev Idx": 11807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564859859.279, "dur": 39.945,
    "args": {
      "Python parent id": 4975, "Python id": 4978, "finished": true, "Ev Idx": 11808
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564859900.435, "dur": 3.414,
    "args": {
      "Python parent id": 4970, "Python id": 4979, "finished": true, "Ev Idx": 11809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564859904.403, "dur": 4.099,
    "args": {
      "Python parent id": 4970, "Python id": 4980, "finished": true, "Ev Idx": 11810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859909.378, "dur": 0.363,
    "args": {
      "Python parent id": 4970, "Python id": 4981, "finished": true, "Ev Idx": 11811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_218", "pid": 537, "tid": 537,
    "ts": 7429564859911.822, "dur": 44.979,
    "args": {
      "Python parent id": 4970, "Python module id": 218, "Python id": 4982, "finished": true, "Ev Idx": 11812
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859912.317, "dur": 44.239,
    "args": {
      "Python parent id": 4982, "Python id": 4983, "finished": true, "Ev Idx": 11813
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859912.592, "dur": 0.258,
    "args": {
      "Python parent id": 4983, "Python id": 4984, "finished": true, "Ev Idx": 11814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564859913.531, "dur": 42.880,
    "args": {
      "Python parent id": 4983, "Python id": 4985, "finished": true, "Ev Idx": 11815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859913.861, "dur": 0.332,
    "args": {
      "Python parent id": 4985, "Python id": 4986, "finished": true, "Ev Idx": 11816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859914.383, "dur": 0.113,
    "args": {
      "Python parent id": 4985, "Python id": 4987, "finished": true, "Ev Idx": 11817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564859914.610, "dur": 41.716,
    "args": {
      "Python parent id": 4985, "Python id": 4988, "finished": true, "Ev Idx": 11818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564859957.350, "dur": 2.766,
    "args": {
      "Python parent id": 4970, "Python id": 4989, "finished": true, "Ev Idx": 11819
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564859960.635, "dur": 2.880,
    "args": {
      "Python parent id": 4970, "Python id": 4990, "finished": true, "Ev Idx": 11820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859964.324, "dur": 0.429,
    "args": {
      "Python parent id": 4970, "Python id": 4991, "finished": true, "Ev Idx": 11821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_219", "pid": 537, "tid": 537,
    "ts": 7429564859966.789, "dur": 40.500,
    "args": {
      "Python parent id": 4970, "Python module id": 219, "Python id": 4992, "finished": true, "Ev Idx": 11822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564859967.230, "dur": 39.813,
    "args": {
      "Python parent id": 4992, "Python id": 4993, "finished": true, "Ev Idx": 11823
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564859967.436, "dur": 0.254,
    "args": {
      "Python parent id": 4993, "Python id": 4994, "finished": true, "Ev Idx": 11824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564859968.354, "dur": 38.402,
    "args": {
      "Python parent id": 4993, "Python id": 4995, "finished": true, "Ev Idx": 11825
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859968.578, "dur": 0.145,
    "args": {
      "Python parent id": 4995, "Python id": 4996, "finished": true, "Ev Idx": 11826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564859968.931, "dur": 0.123,
    "args": {
      "Python parent id": 4995, "Python id": 4997, "finished": true, "Ev Idx": 11827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564859969.163, "dur": 37.549,
    "args": {
      "Python parent id": 4995, "Python id": 4998, "finished": true, "Ev Idx": 11828
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method view of Tensor object at 0x7fe33db6d720>", "pid": 537, "tid": 537,
    "ts": 7429564860007.982, "dur": 2.413,
    "args": {
      "Python parent id": 4970, "Python id": 4999, "finished": true, "Ev Idx": 11829
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564860010.853, "dur": 2.647,
    "args": {
      "Python parent id": 4970, "Python id": 5000, "finished": true, "Ev Idx": 11830
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(133): apply_rotary_pos_emb", "pid": 537, "tid": 537,
    "ts": 7429564860014.714, "dur": 172.260,
    "args": {
      "Python parent id": 4970, "Python id": 5001, "finished": true, "Ev Idx": 11831
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564860014.948, "dur": 3.694,
    "args": {
      "Python parent id": 5001, "Python id": 5002, "finished": true, "Ev Idx": 11832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method unsqueeze of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564860018.931, "dur": 1.803,
    "args": {
      "Python parent id": 5001, "Python id": 5003, "finished": true, "Ev Idx": 11833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564860034.072, "dur": 49.062,
    "args": {
      "Python parent id": 5001, "Python id": 5004, "finished": true, "Ev Idx": 11834
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564860061.462, "dur": 20.521,
    "args": {
      "Python parent id": 5004, "Python id": 5005, "finished": true, "Ev Idx": 11835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(126): rotate_half", "pid": 537, "tid": 537,
    "ts": 7429564860141.703, "dur": 26.825,
    "args": {
      "Python parent id": 5001, "Python id": 5006, "finished": true, "Ev Idx": 11836
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method cat of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564860157.028, "dur": 10.637,
    "args": {
      "Python parent id": 5006, "Python id": 5007, "finished": true, "Ev Idx": 11837
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/cache_utils.py(404): update", "pid": 537, "tid": 537,
    "ts": 7429564860191.087, "dur": 3.988,
    "args": {
      "Python parent id": 4970, "Python id": 5008, "finished": true, "Ev Idx": 11838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564860191.892, "dur": 0.357,
    "args": {
      "Python parent id": 5008, "Python id": 5009, "finished": true, "Ev Idx": 11839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564860192.710, "dur": 0.034,
    "args": {
      "Python parent id": 5008, "Python id": 5010, "finished": true, "Ev Idx": 11840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564860194.001, "dur": 0.252,
    "args": {
      "Python parent id": 5008, "Python id": 5011, "finished": true, "Ev Idx": 11841
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method append of list object at 0x7febfaac4a80>", "pid": 537, "tid": 537,
    "ts": 7429564860194.718, "dur": 0.123,
    "args": {
      "Python parent id": 5008, "Python id": 5012, "finished": true, "Ev Idx": 11842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564860196.379, "dur": 6.561,
    "args": {
      "Python parent id": 4970, "Python id": 5013, "finished": true, "Ev Idx": 11843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(354): _attn_implementation", "pid": 537, "tid": 537,
    "ts": 7429564860199.475, "dur": 3.335,
    "args": {
      "Python parent id": 5013, "Python id": 5014, "finished": true, "Ev Idx": 11844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564860199.742, "dur": 1.356,
    "args": {
      "Python parent id": 5014, "Python id": 5015, "finished": true, "Ev Idx": 11845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564860200.253, "dur": 0.701,
    "args": {
      "Python parent id": 5015, "Python id": 5016, "finished": true, "Ev Idx": 11846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564860201.387, "dur": 0.464,
    "args": {
      "Python parent id": 5014, "Python id": 5017, "finished": true, "Ev Idx": 11847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/configuration_utils.py(207): __getattribute__", "pid": 537, "tid": 537,
    "ts": 7429564860202.147, "dur": 0.592,
    "args": {
      "Python parent id": 5014, "Python id": 5018, "finished": true, "Ev Idx": 11848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(188): eager_attention_forward", "pid": 537, "tid": 537,
    "ts": 7429564860205.756, "dur": 254.971,
    "args": {
      "Python parent id": 4970, "Python id": 5019, "finished": true, "Ev Idx": 11849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564860206.322, "dur": 44.821,
    "args": {
      "Python parent id": 5019, "Python id": 5020, "finished": true, "Ev Idx": 11850
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564860224.117, "dur": 5.097,
    "args": {
      "Python parent id": 5020, "Python id": 5021, "finished": true, "Ev Idx": 11851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564860230.107, "dur": 20.877,
    "args": {
      "Python parent id": 5020, "Python id": 5022, "finished": true, "Ev Idx": 11852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(176): repeat_kv", "pid": 537, "tid": 537,
    "ts": 7429564860251.945, "dur": 31.327,
    "args": {
      "Python parent id": 5019, "Python id": 5023, "finished": true, "Ev Idx": 11853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method expand of Tensor object at 0x7fe33dc62fd0>", "pid": 537, "tid": 537,
    "ts": 7429564860266.400, "dur": 2.612,
    "args": {
      "Python parent id": 5023, "Python id": 5024, "finished": true, "Ev Idx": 11854
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564860269.470, "dur": 13.709,
    "args": {
      "Python parent id": 5023, "Python id": 5025, "finished": true, "Ev Idx": 11855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564860284.185, "dur": 2.931,
    "args": {
      "Python parent id": 5019, "Python id": 5026, "finished": true, "Ev Idx": 11856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564860287.539, "dur": 47.471,
    "args": {
      "Python parent id": 5019, "Python id": 5027, "finished": true, "Ev Idx": 11857
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2103): softmax", "pid": 537, "tid": 537,
    "ts": 7429564860369.228, "dur": 16.491,
    "args": {
      "Python parent id": 5019, "Python id": 5028, "finished": true, "Ev Idx": 11858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564860369.421, "dur": 0.280,
    "args": {
      "Python parent id": 5028, "Python id": 5029, "finished": true, "Ev Idx": 11859
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method softmax of Tensor object at 0x7fe33db85860>", "pid": 537, "tid": 537,
    "ts": 7429564860370.439, "dur": 15.125,
    "args": {
      "Python parent id": 5028, "Python id": 5030, "finished": true, "Ev Idx": 11860
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564860386.725, "dur": 1.651,
    "args": {
      "Python parent id": 5019, "Python id": 5031, "finished": true, "Ev Idx": 11861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(1401): dropout", "pid": 537, "tid": 537,
    "ts": 7429564860389.927, "dur": 9.623,
    "args": {
      "Python parent id": 5019, "Python id": 5032, "finished": true, "Ev Idx": 11862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564860390.005, "dur": 0.085,
    "args": {
      "Python parent id": 5032, "Python id": 5033, "finished": true, "Ev Idx": 11863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_VF.py(27): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860395.596, "dur": 0.919,
    "args": {
      "Python parent id": 5032, "Python id": 5034, "finished": true, "Ev Idx": 11864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564860396.043, "dur": 0.386,
    "args": {
      "Python parent id": 5034, "Python id": 5035, "finished": true, "Ev Idx": 11865
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method dropout of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564860396.897, "dur": 2.561,
    "args": {
      "Python parent id": 5032, "Python id": 5036, "finished": true, "Ev Idx": 11866
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method matmul of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564860400.254, "dur": 41.627,
    "args": {
      "Python parent id": 5019, "Python id": 5037, "finished": true, "Ev Idx": 11867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method transpose of Tensor object at 0x7fe33db6e530>", "pid": 537, "tid": 537,
    "ts": 7429564860442.515, "dur": 2.942,
    "args": {
      "Python parent id": 5019, "Python id": 5038, "finished": true, "Ev Idx": 11868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564860446.210, "dur": 13.374,
    "args": {
      "Python parent id": 5019, "Python id": 5039, "finished": true, "Ev Idx": 11869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method reshape of Tensor object at 0x7fe33dc636b0>", "pid": 537, "tid": 537,
    "ts": 7429564860463.579, "dur": 3.311,
    "args": {
      "Python parent id": 4970, "Python id": 5040, "finished": true, "Ev Idx": 11870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method contiguous of Tensor object at 0x7fe33db86d00>", "pid": 537, "tid": 537,
    "ts": 7429564860467.617, "dur": 0.249,
    "args": {
      "Python parent id": 4970, "Python id": 5041, "finished": true, "Ev Idx": 11871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860468.846, "dur": 0.765,
    "args": {
      "Python parent id": 4970, "Python id": 5042, "finished": true, "Ev Idx": 11872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_220", "pid": 537, "tid": 537,
    "ts": 7429564860473.381, "dur": 41.284,
    "args": {
      "Python parent id": 4970, "Python module id": 220, "Python id": 5043, "finished": true, "Ev Idx": 11873
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564860474.275, "dur": 39.990,
    "args": {
      "Python parent id": 5043, "Python id": 5044, "finished": true, "Ev Idx": 11874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564860474.701, "dur": 1.013,
    "args": {
      "Python parent id": 5044, "Python id": 5045, "finished": true, "Ev Idx": 11875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564860476.790, "dur": 37.280,
    "args": {
      "Python parent id": 5044, "Python id": 5046, "finished": true, "Ev Idx": 11876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860477.260, "dur": 0.255,
    "args": {
      "Python parent id": 5046, "Python id": 5047, "finished": true, "Ev Idx": 11877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860477.770, "dur": 0.156,
    "args": {
      "Python parent id": 5046, "Python id": 5048, "finished": true, "Ev Idx": 11878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564860478.104, "dur": 35.849,
    "args": {
      "Python parent id": 5046, "Python id": 5049, "finished": true, "Ev Idx": 11879
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860528.316, "dur": 0.683,
    "args": {
      "Python parent id": 4954, "Python id": 5050, "finished": true, "Ev Idx": 11880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_63", "pid": 537, "tid": 537,
    "ts": 7429564860531.532, "dur": 77.334,
    "args": {
      "Python parent id": 4954, "Python module id": 63, "Python id": 5051, "finished": true, "Ev Idx": 11881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564860532.113, "dur": 76.488,
    "args": {
      "Python parent id": 5051, "Python id": 5052, "finished": true, "Ev Idx": 11882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564860532.349, "dur": 0.507,
    "args": {
      "Python parent id": 5052, "Python id": 5053, "finished": true, "Ev Idx": 11883
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564860533.620, "dur": 74.152,
    "args": {
      "Python parent id": 5052, "Python id": 5054, "finished": true, "Ev Idx": 11884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564860534.249, "dur": 1.503,
    "args": {
      "Python parent id": 5054, "Python id": 5055, "finished": true, "Ev Idx": 11885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564860536.118, "dur": 13.004,
    "args": {
      "Python parent id": 5054, "Python id": 5056, "finished": true, "Ev Idx": 11886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564860549.516, "dur": 13.641,
    "args": {
      "Python parent id": 5054, "Python id": 5057, "finished": true, "Ev Idx": 11887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564860575.740, "dur": 9.863,
    "args": {
      "Python parent id": 5054, "Python id": 5058, "finished": true, "Ev Idx": 11888
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860595.846, "dur": 0.301,
    "args": {
      "Python parent id": 5054, "Python id": 5059, "finished": true, "Ev Idx": 11889
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564860596.506, "dur": 1.182,
    "args": {
      "Python parent id": 5054, "Python id": 5060, "finished": true, "Ev Idx": 11890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860609.577, "dur": 0.401,
    "args": {
      "Python parent id": 4954, "Python id": 5061, "finished": true, "Ev Idx": 11891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_31", "pid": 537, "tid": 537,
    "ts": 7429564860611.414, "dur": 294.684,
    "args": {
      "Python parent id": 4954, "Python module id": 31, "Python id": 5062, "finished": true, "Ev Idx": 11892
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564860612.059, "dur": 293.865,
    "args": {
      "Python parent id": 5062, "Python id": 5063, "finished": true, "Ev Idx": 11893
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564860612.323, "dur": 0.375,
    "args": {
      "Python parent id": 5063, "Python id": 5064, "finished": true, "Ev Idx": 11894
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(171): forward", "pid": 537, "tid": 537,
    "ts": 7429564860613.890, "dur": 291.899,
    "args": {
      "Python parent id": 5063, "Python id": 5065, "finished": true, "Ev Idx": 11895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860614.162, "dur": 0.361,
    "args": {
      "Python parent id": 5065, "Python id": 5066, "finished": true, "Ev Idx": 11896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860614.835, "dur": 0.325,
    "args": {
      "Python parent id": 5065, "Python id": 5067, "finished": true, "Ev Idx": 11897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860615.454, "dur": 0.203,
    "args": {
      "Python parent id": 5065, "Python id": 5068, "finished": true, "Ev Idx": 11898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_221", "pid": 537, "tid": 537,
    "ts": 7429564860617.192, "dur": 119.548,
    "args": {
      "Python parent id": 5065, "Python module id": 221, "Python id": 5069, "finished": true, "Ev Idx": 11899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564860617.673, "dur": 118.790,
    "args": {
      "Python parent id": 5069, "Python id": 5070, "finished": true, "Ev Idx": 11900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564860617.967, "dur": 0.139,
    "args": {
      "Python parent id": 5070, "Python id": 5071, "finished": true, "Ev Idx": 11901
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564860618.721, "dur": 117.475,
    "args": {
      "Python parent id": 5070, "Python id": 5072, "finished": true, "Ev Idx": 11902
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860618.998, "dur": 0.181,
    "args": {
      "Python parent id": 5072, "Python id": 5073, "finished": true, "Ev Idx": 11903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860619.413, "dur": 0.296,
    "args": {
      "Python parent id": 5072, "Python id": 5074, "finished": true, "Ev Idx": 11904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564860619.836, "dur": 116.245,
    "args": {
      "Python parent id": 5072, "Python id": 5075, "finished": true, "Ev Idx": 11905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_31", "pid": 537, "tid": 537,
    "ts": 7429564860738.796, "dur": 17.313,
    "args": {
      "Python parent id": 5065, "Python module id": 31, "Python id": 5076, "finished": true, "Ev Idx": 11906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564860739.508, "dur": 16.270,
    "args": {
      "Python parent id": 5076, "Python id": 5077, "finished": true, "Ev Idx": 11907
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564860739.746, "dur": 0.539,
    "args": {
      "Python parent id": 5077, "Python id": 5078, "finished": true, "Ev Idx": 11908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/activation.py(431): forward", "pid": 537, "tid": 537,
    "ts": 7429564860741.225, "dur": 14.442,
    "args": {
      "Python parent id": 5077, "Python id": 5079, "finished": true, "Ev Idx": 11909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/functional.py(2358): silu", "pid": 537, "tid": 537,
    "ts": 7429564860742.204, "dur": 13.308,
    "args": {
      "Python parent id": 5079, "Python id": 5080, "finished": true, "Ev Idx": 11910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _has_torch_function_unary>", "pid": 537, "tid": 537,
    "ts": 7429564860742.343, "dur": 0.190,
    "args": {
      "Python parent id": 5080, "Python id": 5081, "finished": true, "Ev Idx": 11911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 537, "tid": 537,
    "ts": 7429564860743.485, "dur": 11.900,
    "args": {
      "Python parent id": 5080, "Python id": 5082, "finished": true, "Ev Idx": 11912
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860757.015, "dur": 0.397,
    "args": {
      "Python parent id": 5065, "Python id": 5083, "finished": true, "Ev Idx": 11913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_222", "pid": 537, "tid": 537,
    "ts": 7429564860759.485, "dur": 83.105,
    "args": {
      "Python parent id": 5065, "Python module id": 222, "Python id": 5084, "finished": true, "Ev Idx": 11914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564860760.001, "dur": 82.319,
    "args": {
      "Python parent id": 5084, "Python id": 5085, "finished": true, "Ev Idx": 11915
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564860760.198, "dur": 0.226,
    "args": {
      "Python parent id": 5085, "Python id": 5086, "finished": true, "Ev Idx": 11916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564860761.115, "dur": 81.020,
    "args": {
      "Python parent id": 5085, "Python id": 5087, "finished": true, "Ev Idx": 11917
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860761.395, "dur": 0.142,
    "args": {
      "Python parent id": 5087, "Python id": 5088, "finished": true, "Ev Idx": 11918
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860761.734, "dur": 0.113,
    "args": {
      "Python parent id": 5087, "Python id": 5089, "finished": true, "Ev Idx": 11919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564860761.981, "dur": 80.035,
    "args": {
      "Python parent id": 5087, "Python id": 5090, "finished": true, "Ev Idx": 11920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_223", "pid": 537, "tid": 537,
    "ts": 7429564860856.005, "dur": 49.475,
    "args": {
      "Python parent id": 5065, "Python module id": 223, "Python id": 5091, "finished": true, "Ev Idx": 11921
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564860856.770, "dur": 48.536,
    "args": {
      "Python parent id": 5091, "Python id": 5092, "finished": true, "Ev Idx": 11922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564860857.003, "dur": 0.486,
    "args": {
      "Python parent id": 5092, "Python id": 5093, "finished": true, "Ev Idx": 11923
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564860858.182, "dur": 46.979,
    "args": {
      "Python parent id": 5092, "Python id": 5094, "finished": true, "Ev Idx": 11924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860858.521, "dur": 0.229,
    "args": {
      "Python parent id": 5094, "Python id": 5095, "finished": true, "Ev Idx": 11925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860859.000, "dur": 0.163,
    "args": {
      "Python parent id": 5094, "Python id": 5096, "finished": true, "Ev Idx": 11926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564860859.271, "dur": 45.794,
    "args": {
      "Python parent id": 5094, "Python id": 5097, "finished": true, "Ev Idx": 11927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860924.281, "dur": 1.055,
    "args": {
      "Python parent id": 64, "Python id": 5098, "finished": true, "Ev Idx": 11928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_64", "pid": 537, "tid": 537,
    "ts": 7429564860927.934, "dur": 77.460,
    "args": {
      "Python parent id": 64, "Python module id": 64, "Python id": 5099, "finished": true, "Ev Idx": 11929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564860928.550, "dur": 76.441,
    "args": {
      "Python parent id": 5099, "Python id": 5100, "finished": true, "Ev Idx": 11930
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564860928.806, "dur": 0.427,
    "args": {
      "Python parent id": 5100, "Python id": 5101, "finished": true, "Ev Idx": 11931
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/models/llama/modeling_llama.py(78): forward", "pid": 537, "tid": 537,
    "ts": 7429564860929.892, "dur": 74.340,
    "args": {
      "Python parent id": 5100, "Python id": 5102, "finished": true, "Ev Idx": 11932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564860930.592, "dur": 1.836,
    "args": {
      "Python parent id": 5102, "Python id": 5103, "finished": true, "Ev Idx": 11933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method pow of Tensor object at 0x7fe33dc7c280>", "pid": 537, "tid": 537,
    "ts": 7429564860932.767, "dur": 11.914,
    "args": {
      "Python parent id": 5102, "Python id": 5104, "finished": true, "Ev Idx": 11934
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method mean of Tensor object at 0x7fe33db6d6d0>", "pid": 537, "tid": 537,
    "ts": 7429564860945.010, "dur": 15.009,
    "args": {
      "Python parent id": 5102, "Python id": 5105, "finished": true, "Ev Idx": 11935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method rsqrt of type object at 0x7fee3795bf40>", "pid": 537, "tid": 537,
    "ts": 7429564860971.524, "dur": 11.726,
    "args": {
      "Python parent id": 5102, "Python id": 5106, "finished": true, "Ev Idx": 11936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564860992.718, "dur": 0.288,
    "args": {
      "Python parent id": 5102, "Python id": 5107, "finished": true, "Ev Idx": 11937
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method to of Tensor object at 0x7fe33dc63610>", "pid": 537, "tid": 537,
    "ts": 7429564860993.392, "dur": 1.210,
    "args": {
      "Python parent id": 5102, "Python id": 5108, "finished": true, "Ev Idx": 11938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<string>(2): __init__", "pid": 537, "tid": 537,
    "ts": 7429564861010.112, "dur": 70.109,
    "args": {
      "Python parent id": 64, "Python id": 5109, "finished": true, "Ev Idx": 11939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(449): __setattr__", "pid": 537, "tid": 537,
    "ts": 7429564861012.142, "dur": 29.246,
    "args": {
      "Python parent id": 5109, "Python id": 5110, "finished": true, "Ev Idx": 11940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method keys of BaseModelOutputWithPast object at 0x7fe33db5ed40>", "pid": 537, "tid": 537,
    "ts": 7429564861038.372, "dur": 0.432,
    "args": {
      "Python parent id": 5110, "Python id": 5111, "finished": true, "Ev Idx": 11941
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(449): __setattr__", "pid": 537, "tid": 537,
    "ts": 7429564861042.084, "dur": 0.933,
    "args": {
      "Python parent id": 5109, "Python id": 5112, "finished": true, "Ev Idx": 11942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method keys of BaseModelOutputWithPast object at 0x7fe33db5ed40>", "pid": 537, "tid": 537,
    "ts": 7429564861042.404, "dur": 0.133,
    "args": {
      "Python parent id": 5112, "Python id": 5113, "finished": true, "Ev Idx": 11943
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(449): __setattr__", "pid": 537, "tid": 537,
    "ts": 7429564861043.451, "dur": 0.665,
    "args": {
      "Python parent id": 5109, "Python id": 5114, "finished": true, "Ev Idx": 11944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method keys of BaseModelOutputWithPast object at 0x7fe33db5ed40>", "pid": 537, "tid": 537,
    "ts": 7429564861043.579, "dur": 0.104,
    "args": {
      "Python parent id": 5114, "Python id": 5115, "finished": true, "Ev Idx": 11945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(449): __setattr__", "pid": 537, "tid": 537,
    "ts": 7429564861044.532, "dur": 0.768,
    "args": {
      "Python parent id": 5109, "Python id": 5116, "finished": true, "Ev Idx": 11946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method keys of BaseModelOutputWithPast object at 0x7fe33db5ed40>", "pid": 537, "tid": 537,
    "ts": 7429564861044.699, "dur": 0.115,
    "args": {
      "Python parent id": 5116, "Python id": 5117, "finished": true, "Ev Idx": 11947
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(374): __post_init__", "pid": 537, "tid": 537,
    "ts": 7429564861049.161, "dur": 30.691,
    "args": {
      "Python parent id": 5109, "Python id": 5118, "finished": true, "Ev Idx": 11948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1278): fields", "pid": 537, "tid": 537,
    "ts": 7429564861050.484, "dur": 7.661,
    "args": {
      "Python parent id": 5118, "Python id": 5119, "finished": true, "Ev Idx": 11949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861051.002, "dur": 0.406,
    "args": {
      "Python parent id": 5119, "Python id": 5120, "finished": true, "Ev Idx": 11950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of dict object at 0x7fe33db43d00>", "pid": 537, "tid": 537,
    "ts": 7429564861052.530, "dur": 0.278,
    "args": {
      "Python parent id": 5119, "Python id": 5121, "finished": true, "Ev Idx": 11951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861055.453, "dur": 0.494,
    "args": {
      "Python parent id": 5119, "Python id": 5122, "finished": true, "Ev Idx": 11952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861056.364, "dur": 0.157,
    "args": {
      "Python parent id": 5119, "Python id": 5123, "finished": true, "Ev Idx": 11953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861056.681, "dur": 0.069,
    "args": {
      "Python parent id": 5119, "Python id": 5124, "finished": true, "Ev Idx": 11954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861057.046, "dur": 0.046,
    "args": {
      "Python parent id": 5119, "Python id": 5125, "finished": true, "Ev Idx": 11955
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861057.273, "dur": 0.100,
    "args": {
      "Python parent id": 5119, "Python id": 5126, "finished": true, "Ev Idx": 11956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564861058.959, "dur": 0.138,
    "args": {
      "Python parent id": 5118, "Python id": 5127, "finished": true, "Ev Idx": 11957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function all>", "pid": 537, "tid": 537,
    "ts": 7429564861061.322, "dur": 2.735,
    "args": {
      "Python parent id": 5118, "Python id": 5128, "finished": true, "Ev Idx": 11958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(384): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861062.351, "dur": 0.170,
    "args": {
      "Python parent id": 5128, "Python id": 5129, "finished": true, "Ev Idx": 11959
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(384): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861062.855, "dur": 0.109,
    "args": {
      "Python parent id": 5128, "Python id": 5130, "finished": true, "Ev Idx": 11960
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(384): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861063.146, "dur": 0.041,
    "args": {
      "Python parent id": 5128, "Python id": 5131, "finished": true, "Ev Idx": 11961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(384): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861063.345, "dur": 0.456,
    "args": {
      "Python parent id": 5128, "Python id": 5132, "finished": true, "Ev Idx": 11962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861064.646, "dur": 0.232,
    "args": {
      "Python parent id": 5118, "Python id": 5133, "finished": true, "Ev Idx": 11963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function all>", "pid": 537, "tid": 537,
    "ts": 7429564861065.844, "dur": 2.071,
    "args": {
      "Python parent id": 5118, "Python id": 5134, "finished": true, "Ev Idx": 11964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(388): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861066.986, "dur": 0.836,
    "args": {
      "Python parent id": 5134, "Python id": 5135, "finished": true, "Ev Idx": 11965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861067.559, "dur": 0.161,
    "args": {
      "Python parent id": 5135, "Python id": 5136, "finished": true, "Ev Idx": 11966
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(388): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861069.195, "dur": 0.741,
    "args": {
      "Python parent id": 5118, "Python id": 5137, "finished": true, "Ev Idx": 11967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861072.451, "dur": 0.082,
    "args": {
      "Python parent id": 5118, "Python id": 5138, "finished": true, "Ev Idx": 11968
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(455): __setitem__", "pid": 537, "tid": 537,
    "ts": 7429564861074.698, "dur": 1.572,
    "args": {
      "Python parent id": 5118, "Python id": 5139, "finished": true, "Ev Idx": 11969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861076.985, "dur": 0.129,
    "args": {
      "Python parent id": 5118, "Python id": 5140, "finished": true, "Ev Idx": 11970
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(455): __setitem__", "pid": 537, "tid": 537,
    "ts": 7429564861077.725, "dur": 0.773,
    "args": {
      "Python parent id": 5118, "Python id": 5141, "finished": true, "Ev Idx": 11971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861078.978, "dur": 0.084,
    "args": {
      "Python parent id": 5118, "Python id": 5142, "finished": true, "Ev Idx": 11972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861079.562, "dur": 0.119,
    "args": {
      "Python parent id": 5118, "Python id": 5143, "finished": true, "Ev Idx": 11973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564861088.955, "dur": 0.149,
    "args": {
      "Python parent id": 48, "Python id": 5144, "finished": true, "Ev Idx": 11974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564861090.766, "dur": 0.980,
    "args": {
      "Python parent id": 48, "Python id": 5145, "finished": true, "Ev Idx": 11975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_224", "pid": 537, "tid": 537,
    "ts": 7429564861111.239, "dur": 55.024,
    "args": {
      "Python parent id": 48, "Python module id": 224, "Python id": 5146, "finished": true, "Ev Idx": 11976
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1755): _call_impl", "pid": 537, "tid": 537,
    "ts": 7429564861112.080, "dur": 53.642,
    "args": {
      "Python parent id": 5146, "Python id": 5147, "finished": true, "Ev Idx": 11977
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _get_tracing_state of PyCapsule object at 0x7fed6a3f1e00>", "pid": 537, "tid": 537,
    "ts": 7429564861112.480, "dur": 0.769,
    "args": {
      "Python parent id": 5147, "Python id": 5148, "finished": true, "Ev Idx": 11978
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/linear.py(124): forward", "pid": 537, "tid": 537,
    "ts": 7429564861114.239, "dur": 51.207,
    "args": {
      "Python parent id": 5147, "Python id": 5149, "finished": true, "Ev Idx": 11979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564861114.633, "dur": 0.243,
    "args": {
      "Python parent id": 5149, "Python id": 5150, "finished": true, "Ev Idx": 11980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/nn/modules/module.py(1927): __getattr__", "pid": 537, "tid": 537,
    "ts": 7429564861115.162, "dur": 0.175,
    "args": {
      "Python parent id": 5149, "Python id": 5151, "finished": true, "Ev Idx": 11981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 537, "tid": 537,
    "ts": 7429564861115.553, "dur": 49.784,
    "args": {
      "Python parent id": 5149, "Python id": 5152, "finished": true, "Ev Idx": 11982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<string>(2): __init__", "pid": 537, "tid": 537,
    "ts": 7429564861170.440, "dur": 24.227,
    "args": {
      "Python parent id": 48, "Python id": 5153, "finished": true, "Ev Idx": 11983
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(449): __setattr__", "pid": 537, "tid": 537,
    "ts": 7429564861170.889, "dur": 1.275,
    "args": {
      "Python parent id": 5153, "Python id": 5154, "finished": true, "Ev Idx": 11984
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method keys of BaseModelOutputWithPast object at 0x7fe33db5ed40>", "pid": 537, "tid": 537,
    "ts": 7429564861171.141, "dur": 0.187,
    "args": {
      "Python parent id": 5154, "Python id": 5155, "finished": true, "Ev Idx": 11985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(449): __setattr__", "pid": 537, "tid": 537,
    "ts": 7429564861172.583, "dur": 0.796,
    "args": {
      "Python parent id": 5153, "Python id": 5156, "finished": true, "Ev Idx": 11986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method keys of BaseModelOutputWithPast object at 0x7fe33db5ed40>", "pid": 537, "tid": 537,
    "ts": 7429564861172.831, "dur": 0.097,
    "args": {
      "Python parent id": 5156, "Python id": 5157, "finished": true, "Ev Idx": 11987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(449): __setattr__", "pid": 537, "tid": 537,
    "ts": 7429564861173.656, "dur": 0.531,
    "args": {
      "Python parent id": 5153, "Python id": 5158, "finished": true, "Ev Idx": 11988
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method keys of BaseModelOutputWithPast object at 0x7fe33db5ed40>", "pid": 537, "tid": 537,
    "ts": 7429564861173.771, "dur": 0.110,
    "args": {
      "Python parent id": 5158, "Python id": 5159, "finished": true, "Ev Idx": 11989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(449): __setattr__", "pid": 537, "tid": 537,
    "ts": 7429564861174.435, "dur": 0.531,
    "args": {
      "Python parent id": 5153, "Python id": 5160, "finished": true, "Ev Idx": 11990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method keys of BaseModelOutputWithPast object at 0x7fe33db5ed40>", "pid": 537, "tid": 537,
    "ts": 7429564861174.542, "dur": 0.096,
    "args": {
      "Python parent id": 5160, "Python id": 5161, "finished": true, "Ev Idx": 11991
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(449): __setattr__", "pid": 537, "tid": 537,
    "ts": 7429564861175.289, "dur": 0.823,
    "args": {
      "Python parent id": 5153, "Python id": 5162, "finished": true, "Ev Idx": 11992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method keys of BaseModelOutputWithPast object at 0x7fe33db5ed40>", "pid": 537, "tid": 537,
    "ts": 7429564861175.410, "dur": 0.060,
    "args": {
      "Python parent id": 5162, "Python id": 5163, "finished": true, "Ev Idx": 11993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(374): __post_init__", "pid": 537, "tid": 537,
    "ts": 7429564861177.164, "dur": 17.254,
    "args": {
      "Python parent id": 5153, "Python id": 5164, "finished": true, "Ev Idx": 11994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1278): fields", "pid": 537, "tid": 537,
    "ts": 7429564861177.575, "dur": 4.476,
    "args": {
      "Python parent id": 5164, "Python id": 5165, "finished": true, "Ev Idx": 11995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861177.876, "dur": 0.274,
    "args": {
      "Python parent id": 5165, "Python id": 5166, "finished": true, "Ev Idx": 11996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method values of dict object at 0x7fe33db43d00>", "pid": 537, "tid": 537,
    "ts": 7429564861178.627, "dur": 0.137,
    "args": {
      "Python parent id": 5165, "Python id": 5167, "finished": true, "Ev Idx": 11997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861179.592, "dur": 0.368,
    "args": {
      "Python parent id": 5165, "Python id": 5168, "finished": true, "Ev Idx": 11998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861180.184, "dur": 0.190,
    "args": {
      "Python parent id": 5165, "Python id": 5169, "finished": true, "Ev Idx": 11999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861180.511, "dur": 0.081,
    "args": {
      "Python parent id": 5165, "Python id": 5170, "finished": true, "Ev Idx": 12000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861180.846, "dur": 0.049,
    "args": {
      "Python parent id": 5165, "Python id": 5171, "finished": true, "Ev Idx": 12001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861181.299, "dur": 0.045,
    "args": {
      "Python parent id": 5165, "Python id": 5172, "finished": true, "Ev Idx": 12002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861181.535, "dur": 0.105,
    "args": {
      "Python parent id": 5165, "Python id": 5173, "finished": true, "Ev Idx": 12003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 537, "tid": 537,
    "ts": 7429564861182.297, "dur": 0.095,
    "args": {
      "Python parent id": 5164, "Python id": 5174, "finished": true, "Ev Idx": 12004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function all>", "pid": 537, "tid": 537,
    "ts": 7429564861182.995, "dur": 1.700,
    "args": {
      "Python parent id": 5164, "Python id": 5175, "finished": true, "Ev Idx": 12005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(384): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861183.325, "dur": 0.171,
    "args": {
      "Python parent id": 5175, "Python id": 5176, "finished": true, "Ev Idx": 12006
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(384): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861183.736, "dur": 0.071,
    "args": {
      "Python parent id": 5175, "Python id": 5177, "finished": true, "Ev Idx": 12007
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(384): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861183.981, "dur": 0.055,
    "args": {
      "Python parent id": 5175, "Python id": 5178, "finished": true, "Ev Idx": 12008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(384): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861184.190, "dur": 0.040,
    "args": {
      "Python parent id": 5175, "Python id": 5179, "finished": true, "Ev Idx": 12009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(384): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861184.384, "dur": 0.135,
    "args": {
      "Python parent id": 5175, "Python id": 5180, "finished": true, "Ev Idx": 12010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861185.070, "dur": 0.136,
    "args": {
      "Python parent id": 5164, "Python id": 5181, "finished": true, "Ev Idx": 12011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function all>", "pid": 537, "tid": 537,
    "ts": 7429564861185.875, "dur": 0.968,
    "args": {
      "Python parent id": 5164, "Python id": 5182, "finished": true, "Ev Idx": 12012
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(388): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861186.243, "dur": 0.527,
    "args": {
      "Python parent id": 5182, "Python id": 5183, "finished": true, "Ev Idx": 12013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861186.517, "dur": 0.138,
    "args": {
      "Python parent id": 5183, "Python id": 5184, "finished": true, "Ev Idx": 12014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(388): <genexpr>", "pid": 537, "tid": 537,
    "ts": 7429564861187.487, "dur": 0.267,
    "args": {
      "Python parent id": 5164, "Python id": 5185, "finished": true, "Ev Idx": 12015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861188.889, "dur": 0.097,
    "args": {
      "Python parent id": 5164, "Python id": 5186, "finished": true, "Ev Idx": 12016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861189.599, "dur": 0.085,
    "args": {
      "Python parent id": 5164, "Python id": 5187, "finished": true, "Ev Idx": 12017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(455): __setitem__", "pid": 537, "tid": 537,
    "ts": 7429564861190.427, "dur": 0.895,
    "args": {
      "Python parent id": 5164, "Python id": 5188, "finished": true, "Ev Idx": 12018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861191.850, "dur": 0.088,
    "args": {
      "Python parent id": 5164, "Python id": 5189, "finished": true, "Ev Idx": 12019
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "transformers/utils/generic.py(455): __setitem__", "pid": 537, "tid": 537,
    "ts": 7429564861192.566, "dur": 0.596,
    "args": {
      "Python parent id": 5164, "Python id": 5190, "finished": true, "Ev Idx": 12020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861193.630, "dur": 0.076,
    "args": {
      "Python parent id": 5164, "Python id": 5191, "finished": true, "Ev Idx": 12021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861194.182, "dur": 0.089,
    "args": {
      "Python parent id": 5164, "Python id": 5192, "finished": true, "Ev Idx": 12022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(803): __exit__", "pid": 537, "tid": 537,
    "ts": 7429564861276.323, "dur": 596.592,
    "args": {
      "Python parent id": 6, "Python id": 5193, "finished": true, "Ev Idx": 12023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(817): stop", "pid": 537, "tid": 537,
    "ts": 7429564861278.531, "dur": 594.384,
    "args": {
      "Python parent id": 5193, "Python id": 5194, "finished": true, "Ev Idx": 12024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(862): _transit_action", "pid": 537, "tid": 537,
    "ts": 7429564861292.215, "dur": 580.700,
    "args": {
      "Python parent id": 5194, "Python id": 5195, "finished": true, "Ev Idx": 12025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method get of dict object at 0x7fe33db65600>", "pid": 537, "tid": 537,
    "ts": 7429564861295.637, "dur": 5.744,
    "args": {
      "Python parent id": 5195, "Python id": 5196, "finished": true, "Ev Idx": 12026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1261): __hash__", "pid": 537, "tid": 537,
    "ts": 7429564861298.484, "dur": 1.793,
    "args": {
      "Python parent id": 5196, "Python id": 5197, "finished": true, "Ev Idx": 12027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hash>", "pid": 537, "tid": 537,
    "ts": 7429564861300.032, "dur": 0.149,
    "args": {
      "Python parent id": 5197, "Python id": 5198, "finished": true, "Ev Idx": 12028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/profiler/profiler.py(239): stop_trace", "pid": 537, "tid": 537,
    "ts": 7429564861303.010, "dur": 569.905,
    "args": {
      "Python parent id": 5195, "Python id": 5199, "finished": true, "Ev Idx": 12029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/autograd/profiler.py(375): __exit__", "pid": 537, "tid": 537,
    "ts": 7429564861307.648, "dur": 565.267,
    "args": {
      "Python parent id": 5199, "Python id": 5200, "finished": true, "Ev Idx": 12030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564861309.328, "dur": 0.704,
    "args": {
      "Python parent id": 5200, "Python id": 5201, "finished": true, "Ev Idx": 12031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 537, "tid": 537,
    "ts": 7429564861314.961, "dur": 0.203,
    "args": {
      "Python parent id": 5200, "Python id": 5202, "finished": true, "Ev Idx": 12032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564861316.046, "dur": 0.838,
    "args": {
      "Python parent id": 5200, "Python id": 5203, "finished": true, "Ev Idx": 12033
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1030): synchronize", "pid": 537, "tid": 537,
    "ts": 7429564861318.861, "dur": 554.054,
    "args": {
      "Python parent id": 5200, "Python id": 5204, "finished": true, "Ev Idx": 12034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 537, "tid": 537,
    "ts": 7429564861323.039, "dur": 6.087,
    "args": {
      "Python parent id": 5204, "Python id": 5205, "finished": true, "Ev Idx": 12035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 537, "tid": 537,
    "ts": 7429564861324.699, "dur": 4.082,
    "args": {
      "Python parent id": 5205, "Python id": 5206, "finished": true, "Ev Idx": 12036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 537, "tid": 537,
    "ts": 7429564861328.195, "dur": 0.426,
    "args": {
      "Python parent id": 5206, "Python id": 5207, "finished": true, "Ev Idx": 12037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(490): __init__", "pid": 537, "tid": 537,
    "ts": 7429564861333.044, "dur": 55.584,
    "args": {
      "Python parent id": 5204, "Python id": 5208, "finished": true, "Ev Idx": 12038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/_utils.py(9): _get_device_index", "pid": 537, "tid": 537,
    "ts": 7429564861336.044, "dur": 51.851,
    "args": {
      "Python parent id": 5208, "Python id": 5209, "finished": true, "Ev Idx": 12039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564861337.228, "dur": 0.116,
    "args": {
      "Python parent id": 5209, "Python id": 5210, "finished": true, "Ev Idx": 12040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564861337.865, "dur": 0.060,
    "args": {
      "Python parent id": 5209, "Python id": 5211, "finished": true, "Ev Idx": 12041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564861338.559, "dur": 0.046,
    "args": {
      "Python parent id": 5209, "Python id": 5212, "finished": true, "Ev Idx": 12042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 537, "tid": 537,
    "ts": 7429564861340.931, "dur": 0.089,
    "args": {
      "Python parent id": 5209, "Python id": 5213, "finished": true, "Ev Idx": 12043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564861342.040, "dur": 0.165,
    "args": {
      "Python parent id": 5209, "Python id": 5214, "finished": true, "Ev Idx": 12044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(812): _get_device_index", "pid": 537, "tid": 537,
    "ts": 7429564861345.232, "dur": 42.512,
    "args": {
      "Python parent id": 5209, "Python id": 5215, "finished": true, "Ev Idx": 12045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564861346.096, "dur": 0.087,
    "args": {
      "Python parent id": 5215, "Python id": 5216, "finished": true, "Ev Idx": 12046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564861346.812, "dur": 0.064,
    "args": {
      "Python parent id": 5215, "Python id": 5217, "finished": true, "Ev Idx": 12047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564861347.665, "dur": 0.078,
    "args": {
      "Python parent id": 5215, "Python id": 5218, "finished": true, "Ev Idx": 12048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_jit_internal.py(101): is_scripting", "pid": 537, "tid": 537,
    "ts": 7429564861348.481, "dur": 0.032,
    "args": {
      "Python parent id": 5215, "Python id": 5219, "finished": true, "Ev Idx": 12049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(786): _get_current_device_index", "pid": 537, "tid": 537,
    "ts": 7429564861350.657, "dur": 36.903,
    "args": {
      "Python parent id": 5215, "Python id": 5220, "finished": true, "Ev Idx": 12050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(770): _get_device_attr", "pid": 537, "tid": 537,
    "ts": 7429564861354.077, "dur": 33.340,
    "args": {
      "Python parent id": 5220, "Python id": 5221, "finished": true, "Ev Idx": 12051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(753): _get_available_device_type", "pid": 537, "tid": 537,
    "ts": 7429564861356.709, "dur": 21.152,
    "args": {
      "Python parent id": 5221, "Python id": 5222, "finished": true, "Ev Idx": 12052
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(161): is_available", "pid": 537, "tid": 537,
    "ts": 7429564861358.510, "dur": 19.075,
    "args": {
      "Python parent id": 5222, "Python id": 5223, "finished": true, "Ev Idx": 12053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(152): _is_compiled", "pid": 537, "tid": 537,
    "ts": 7429564861359.718, "dur": 1.409,
    "args": {
      "Python parent id": 5223, "Python id": 5224, "finished": true, "Ev Idx": 12054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 537, "tid": 537,
    "ts": 7429564861360.461, "dur": 0.593,
    "args": {
      "Python parent id": 5224, "Python id": 5225, "finished": true, "Ev Idx": 12055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(157): _nvml_based_avail", "pid": 537, "tid": 537,
    "ts": 7429564861362.193, "dur": 13.513,
    "args": {
      "Python parent id": 5223, "Python id": 5226, "finished": true, "Ev Idx": 12056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 537, "tid": 537,
    "ts": 7429564861364.026, "dur": 11.292,
    "args": {
      "Python parent id": 5226, "Python id": 5227, "finished": true, "Ev Idx": 12057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen _collections_abc>(804): get", "pid": 537, "tid": 537,
    "ts": 7429564861365.469, "dur": 9.696,
    "args": {
      "Python parent id": 5227, "Python id": 5228, "finished": true, "Ev Idx": 12058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 537, "tid": 537,
    "ts": 7429564861366.901, "dur": 7.285,
    "args": {
      "Python parent id": 5228, "Python id": 5229, "finished": true, "Ev Idx": 12059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 537, "tid": 537,
    "ts": 7429564861368.573, "dur": 3.008,
    "args": {
      "Python parent id": 5229, "Python id": 5230, "finished": true, "Ev Idx": 12060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 537, "tid": 537,
    "ts": 7429564861368.867, "dur": 0.061,
    "args": {
      "Python parent id": 5230, "Python id": 5231, "finished": true, "Ev Idx": 12061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method encode of str object at 0x7feca05c57f0>", "pid": 537, "tid": 537,
    "ts": 7429564861370.872, "dur": 0.576,
    "args": {
      "Python parent id": 5230, "Python id": 5232, "finished": true, "Ev Idx": 12062
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDeviceCount>", "pid": 537, "tid": 537,
    "ts": 7429564861376.937, "dur": 0.492,
    "args": {
      "Python parent id": 5223, "Python id": 5233, "finished": true, "Ev Idx": 12063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method lower of str object at 0x7fee3842bb70>", "pid": 537, "tid": 537,
    "ts": 7429564861380.123, "dur": 0.471,
    "args": {
      "Python parent id": 5221, "Python id": 5234, "finished": true, "Ev Idx": 12064
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/_utils.py(788): <lambda>", "pid": 537, "tid": 537,
    "ts": 7429564861381.758, "dur": 5.548,
    "args": {
      "Python parent id": 5221, "Python id": 5235, "finished": true, "Ev Idx": 12065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(1024): current_device", "pid": 537, "tid": 537,
    "ts": 7429564861383.406, "dur": 3.825,
    "args": {
      "Python parent id": 5235, "Python id": 5236, "finished": true, "Ev Idx": 12066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(342): _lazy_init", "pid": 537, "tid": 537,
    "ts": 7429564861383.852, "dur": 1.022,
    "args": {
      "Python parent id": 5236, "Python id": 5237, "finished": true, "Ev Idx": 12067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(294): is_initialized", "pid": 537, "tid": 537,
    "ts": 7429564861384.305, "dur": 0.470,
    "args": {
      "Python parent id": 5237, "Python id": 5238, "finished": true, "Ev Idx": 12068
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_isInBadFork>", "pid": 537, "tid": 537,
    "ts": 7429564861384.622, "dur": 0.100,
    "args": {
      "Python parent id": 5238, "Python id": 5239, "finished": true, "Ev Idx": 12069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_getDevice>", "pid": 537, "tid": 537,
    "ts": 7429564861386.212, "dur": 0.953,
    "args": {
      "Python parent id": 5236, "Python id": 5240, "finished": true, "Ev Idx": 12070
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(494): __enter__", "pid": 537, "tid": 537,
    "ts": 7429564861390.304, "dur": 2.534,
    "args": {
      "Python parent id": 5204, "Python id": 5241, "finished": true, "Ev Idx": 12071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_exchangeDevice>", "pid": 537, "tid": 537,
    "ts": 7429564861391.774, "dur": 0.906,
    "args": {
      "Python parent id": 5241, "Python id": 5242, "finished": true, "Ev Idx": 12072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_synchronize>", "pid": 537, "tid": 537,
    "ts": 7429564861394.558, "dur": 464.143,
    "args": {
      "Python parent id": 5204, "Python id": 5243, "finished": true, "Ev Idx": 12073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "torch/cuda/__init__.py(497): __exit__", "pid": 537, "tid": 537,
    "ts": 7429564861860.099, "dur": 12.816,
    "args": {
      "Python parent id": 5204, "Python id": 5244, "finished": true, "Ev Idx": 12074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function _cuda_maybeExchangeDevice>", "pid": 537, "tid": 537,
    "ts": 7429564861861.858, "dur": 0.898,
    "args": {
      "Python parent id": 5244, "Python id": 5245, "finished": true, "Ev Idx": 12075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function perf_counter_ns>", "pid": 537, "tid": 537,
    "ts": 7429564861864.996, "dur": 0.527,
    "args": {
      "Python parent id": 5244, "Python id": 5246, "finished": true, "Ev Idx": 12076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in method _disable_profiler of PyCapsule object at 0x7febfce6f1b0>", "pid": 537, "tid": 537,
    "ts": 7429564861867.291, "dur": 5.624,
    "args": {
      "Python parent id": 5244, "Python id": 5247, "finished": true, "Ev Idx": 12077
    }
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564820856.478, "dur": 14.046,
    "args": {
      "External id": 4,          "kernel": "void at::native::(anonymous namespace)::indexSelectSmallIndex<float, long, unsigned int, 2, 2, -2>(at::cuda::detail::TensorInfo<float, unsigned int>, at::cuda::detail::TensorInfo<float const, unsigned int>, at::cuda::detail::TensorInfo<long const, unsigned int>, int, int, unsigned int, long)", "cid": 107, "correlation": 1,      "grid": [8, 1, 1],      "block": [512, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1, "pid": 537, "tid": 537, "ts": 7429564820856.478,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564820912.792, "dur": 7.009,
    "args": {
      "External id": 10,          "kernel": "void (anonymous namespace)::elementwise_kernel_with_index<int, at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}>(int, at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, function_traits<at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}>::result_type*)", "cid": 107, "correlation": 2,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2, "pid": 537, "tid": 537, "ts": 7429564820912.792,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564820988.857, "dur": 6.512,
    "args": {
      "External id": 16,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int, at::native::FillFunctor<float>, std::array<char*, 1ul>)", "cid": 107, "correlation": 3,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 3, "pid": 537, "tid": 537, "ts": 7429564820988.857,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821006.006, "dur": 7.045,
    "args": {
      "External id": 17,          "kernel": "void at::native::triu_tril_kernel<float, int, true, 2, false>(at::cuda::detail::TensorInfo<float, int>, at::cuda::detail::TensorInfo<float const, int>, long, long, int)", "cid": 107, "correlation": 4,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 4, "pid": 537, "tid": 537, "ts": 7429564821006.006,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821022.684, "dur": 4.665,
    "args": {
      "External id": 20,          "kernel": "void (anonymous namespace)::elementwise_kernel_with_index<int, at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}>(int, at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, function_traits<at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}>::result_type*)", "cid": 107, "correlation": 5,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 5, "pid": 537, "tid": 537, "ts": 7429564821022.684,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821044.907, "dur": 5.797,
    "args": {
      "External id": 24,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast<at::native::(anonymous namespace)::CompareFunctor<long> >(at::TensorIteratorBase&, at::native::(anonymous namespace)::CompareFunctor<long> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::(anonymous namespace)::CompareFunctor<long> >(at::TensorIteratorBase&, at::native::(anonymous namespace)::CompareFunctor<long> const&)::{lambda(int)#1})", "cid": 107, "correlation": 6,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 6, "pid": 537, "tid": 537, "ts": 7429564821044.907,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821061.382, "dur": 6.897,
    "args": {
      "External id": 25,          "kernel": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 7,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 7, "pid": 537, "tid": 537, "ts": 7429564821061.382,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 537, "tid": 537,
    "ts": 7429564821122.846, "dur": 10.201,
    "args": {
      "External id": 38,      "cid": 134, "correlation": 8, "src": "0x7fe2f1bfbc00", "dst": "0x7fe2f1bfb800", "bytes": "676", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 8, "pid": 537, "tid": 537, "ts": 7429564821122.846,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821167.535, "dur": 6.801,
    "args": {
      "External id": 54,          "kernel": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "cid": 107, "correlation": 9,      "grid": [1, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 9, "pid": 537, "tid": 537, "ts": 7429564821167.535,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821187.936, "dur": 5.446,
    "args": {
      "External id": 55,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, bool, at::native::(anonymous namespace)::CompareEqFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, bool, at::native::(anonymous namespace)::CompareEqFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 10,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 10, "pid": 537, "tid": 537, "ts": 7429564821187.936,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 537, "tid": 537,
    "ts": 7429564821218.540, "dur": 5.506,
    "args": {
      "External id": 66,      "cid": 134, "correlation": 11, "src": "0x7fe2f1bfb800", "dst": "0x7fe2f1bfbc00", "bytes": "676", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 11, "pid": 537, "tid": 537, "ts": 7429564821218.540,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821233.447, "dur": 4.464,
    "args": {
      "External id": 67,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::masked_fill_kernel(at::TensorIterator&, c10::Scalar const&)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float, bool)#1}, std::array<char*, 3ul> >(int, at::native::(anonymous namespace)::masked_fill_kernel(at::TensorIterator&, c10::Scalar const&)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float, bool)#1}, std::array<char*, 3ul>)", "cid": 107, "correlation": 12,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 12, "pid": 537, "tid": 537, "ts": 7429564821233.447,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 537, "tid": 537,
    "ts": 7429564821251.456, "dur": 3.754,
    "args": {
      "External id": 74,      "cid": 134, "correlation": 13, "src": "0x7fe2f1bfbc00", "dst": "0x7fe2f1bfb800", "bytes": "676", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 13, "pid": 537, "tid": 537, "ts": 7429564821251.456,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821336.045, "dur": 5.451,
    "args": {
      "External id": 94,          "kernel": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 14,      "grid": [1, 1, 1],      "block": [512, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 14, "pid": 537, "tid": 537, "ts": 7429564821336.045,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564821441.078, "dur": 0.669,
    "args": {
      "External id": 106,      "cid": 370, "correlation": 15
    }
  },
  {
    "ph": "f", "id": 15, "pid": 537, "tid": 537, "ts": 7429564821441.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564821451.488, "dur": 0.184,
    "args": {
      "External id": 106,      "cid": 370, "correlation": 16
    }
  },
  {
    "ph": "f", "id": 16, "pid": 537, "tid": 537, "ts": 7429564821451.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564821457.136, "dur": 0.526,
    "args": {
      "External id": 106,      "cid": 370, "correlation": 17
    }
  },
  {
    "ph": "f", "id": 17, "pid": 537, "tid": 537, "ts": 7429564821457.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821465.687, "dur": 7.478,
    "args": {
      "External id": 106,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT16x64x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG16_16_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 18,      "grid": [256, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 18, "pid": 537, "tid": 537, "ts": 7429564821465.687,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821500.569, "dur": 7.160,
    "args": {
      "External id": 111,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 3, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 19,      "grid": [1, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 19, "pid": 537, "tid": 537, "ts": 7429564821500.569,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821518.299, "dur": 6.957,
    "args": {
      "External id": 112,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::cos_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::cos_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 20,      "grid": [2, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 20, "pid": 537, "tid": 537, "ts": 7429564821518.299,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821537.523, "dur": 4.595,
    "args": {
      "External id": 113,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 21,      "grid": [2, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 21, "pid": 537, "tid": 537, "ts": 7429564821537.523,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821550.654, "dur": 6.077,
    "args": {
      "External id": 114,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::sin_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::sin_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 22,      "grid": [2, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 22, "pid": 537, "tid": 537, "ts": 7429564821550.654,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821562.019, "dur": 3.175,
    "args": {
      "External id": 115,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 23,      "grid": [2, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 23, "pid": 537, "tid": 537, "ts": 7429564821562.019,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821780.095, "dur": 7.468,
    "args": {
      "External id": 119,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 24,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 24, "pid": 537, "tid": 537, "ts": 7429564821780.095,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821809.948, "dur": 6.670,
    "args": {
      "External id": 122,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 25,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 25, "pid": 537, "tid": 537, "ts": 7429564821809.948,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821829.188, "dur": 6.448,
    "args": {
      "External id": 123,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 26,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 26, "pid": 537, "tid": 537, "ts": 7429564821829.188,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821843.807, "dur": 5.580,
    "args": {
      "External id": 124,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 27,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 27, "pid": 537, "tid": 537, "ts": 7429564821843.807,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821859.518, "dur": 4.866,
    "args": {
      "External id": 125,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 28,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 28, "pid": 537, "tid": 537, "ts": 7429564821859.518,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821874.286, "dur": 3.501,
    "args": {
      "External id": 127,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 29,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 29, "pid": 537, "tid": 537, "ts": 7429564821874.286,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564821958.363, "dur": 0.480,
    "args": {
      "External id": 135,      "cid": 370, "correlation": 30
    }
  },
  {
    "ph": "f", "id": 30, "pid": 537, "tid": 537, "ts": 7429564821958.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564821965.797, "dur": 0.155,
    "args": {
      "External id": 135,      "cid": 370, "correlation": 31
    }
  },
  {
    "ph": "f", "id": 31, "pid": 537, "tid": 537, "ts": 7429564821965.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564821970.375, "dur": 0.820,
    "args": {
      "External id": 135,      "cid": 370, "correlation": 32
    }
  },
  {
    "ph": "f", "id": 32, "pid": 537, "tid": 537, "ts": 7429564821970.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564821976.334, "dur": 8.133,
    "args": {
      "External id": 135,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 33,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 33, "pid": 537, "tid": 537, "ts": 7429564821976.334,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822038.654, "dur": 0.322,
    "args": {
      "External id": 148,      "cid": 370, "correlation": 34
    }
  },
  {
    "ph": "f", "id": 34, "pid": 537, "tid": 537, "ts": 7429564822038.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822042.379, "dur": 0.161,
    "args": {
      "External id": 148,      "cid": 370, "correlation": 35
    }
  },
  {
    "ph": "f", "id": 35, "pid": 537, "tid": 537, "ts": 7429564822042.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822045.131, "dur": 0.174,
    "args": {
      "External id": 148,      "cid": 370, "correlation": 36
    }
  },
  {
    "ph": "f", "id": 36, "pid": 537, "tid": 537, "ts": 7429564822045.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822054.475, "dur": 7.167,
    "args": {
      "External id": 148,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 37,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 37, "pid": 537, "tid": 537, "ts": 7429564822054.475,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822062.216, "dur": 5.383,
    "args": {
      "External id": 148,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 38,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 38, "pid": 537, "tid": 537, "ts": 7429564822062.216,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822144.006, "dur": 0.328,
    "args": {
      "External id": 161,      "cid": 370, "correlation": 39
    }
  },
  {
    "ph": "f", "id": 39, "pid": 537, "tid": 537, "ts": 7429564822144.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822147.523, "dur": 0.165,
    "args": {
      "External id": 161,      "cid": 370, "correlation": 40
    }
  },
  {
    "ph": "f", "id": 40, "pid": 537, "tid": 537, "ts": 7429564822147.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822150.159, "dur": 0.147,
    "args": {
      "External id": 161,      "cid": 370, "correlation": 41
    }
  },
  {
    "ph": "f", "id": 41, "pid": 537, "tid": 537, "ts": 7429564822150.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822155.592, "dur": 5.304,
    "args": {
      "External id": 161,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 42,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 42, "pid": 537, "tid": 537, "ts": 7429564822155.592,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822161.284, "dur": 4.782,
    "args": {
      "External id": 161,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 43,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 43, "pid": 537, "tid": 537, "ts": 7429564822161.284,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822196.295, "dur": 3.682,
    "args": {
      "External id": 171,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 44,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 44, "pid": 537, "tid": 537, "ts": 7429564822196.295,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822225.997, "dur": 6.232,
    "args": {
      "External id": 176,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 45,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 45, "pid": 537, "tid": 537, "ts": 7429564822225.997,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822247.021, "dur": 7.608,
    "args": {
      "External id": 177,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 46,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 46, "pid": 537, "tid": 537, "ts": 7429564822247.021,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822263.303, "dur": 4.553,
    "args": {
      "External id": 178,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 47,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 47, "pid": 537, "tid": 537, "ts": 7429564822263.303,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822275.813, "dur": 5.268,
    "args": {
      "External id": 179,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 48,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 48, "pid": 537, "tid": 537, "ts": 7429564822275.813,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822287.989, "dur": 3.925,
    "args": {
      "External id": 180,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 49,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 49, "pid": 537, "tid": 537, "ts": 7429564822287.989,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822306.709, "dur": 3.172,
    "args": {
      "External id": 185,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 50,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 50, "pid": 537, "tid": 537, "ts": 7429564822306.709,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822316.715, "dur": 4.779,
    "args": {
      "External id": 186,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 51,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 51, "pid": 537, "tid": 537, "ts": 7429564822316.715,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822328.290, "dur": 3.381,
    "args": {
      "External id": 187,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 52,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 52, "pid": 537, "tid": 537, "ts": 7429564822328.290,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822336.270, "dur": 4.930,
    "args": {
      "External id": 188,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 53,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 53, "pid": 537, "tid": 537, "ts": 7429564822336.270,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822416.638, "dur": 5.825,
    "args": {
      "External id": 205,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 54,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 54, "pid": 537, "tid": 537, "ts": 7429564822416.638,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822453.714, "dur": 4.190,
    "args": {
      "External id": 223,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 55,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 55, "pid": 537, "tid": 537, "ts": 7429564822453.714,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822503.145, "dur": 0.327,
    "args": {
      "External id": 236,      "cid": 370, "correlation": 56
    }
  },
  {
    "ph": "f", "id": 56, "pid": 537, "tid": 537, "ts": 7429564822503.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822507.506, "dur": 0.161,
    "args": {
      "External id": 236,      "cid": 370, "correlation": 57
    }
  },
  {
    "ph": "f", "id": 57, "pid": 537, "tid": 537, "ts": 7429564822507.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822510.289, "dur": 0.161,
    "args": {
      "External id": 236,      "cid": 370, "correlation": 58
    }
  },
  {
    "ph": "f", "id": 58, "pid": 537, "tid": 537, "ts": 7429564822510.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822512.943, "dur": 4.785,
    "args": {
      "External id": 236,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 59,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 59, "pid": 537, "tid": 537, "ts": 7429564822512.943,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822532.330, "dur": 5.012,
    "args": {
      "External id": 239,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 60,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 60, "pid": 537, "tid": 537, "ts": 7429564822532.330,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822553.295, "dur": 3.733,
    "args": {
      "External id": 246,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 61,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 61, "pid": 537, "tid": 537, "ts": 7429564822553.295,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822574.513, "dur": 6.763,
    "args": {
      "External id": 249,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 62,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 62, "pid": 537, "tid": 537, "ts": 7429564822574.513,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822629.745, "dur": 0.201,
    "args": {
      "External id": 261,      "cid": 370, "correlation": 63
    }
  },
  {
    "ph": "f", "id": 63, "pid": 537, "tid": 537, "ts": 7429564822629.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822633.544, "dur": 0.172,
    "args": {
      "External id": 261,      "cid": 370, "correlation": 64
    }
  },
  {
    "ph": "f", "id": 64, "pid": 537, "tid": 537, "ts": 7429564822633.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822636.782, "dur": 0.173,
    "args": {
      "External id": 261,      "cid": 370, "correlation": 65
    }
  },
  {
    "ph": "f", "id": 65, "pid": 537, "tid": 537, "ts": 7429564822636.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822640.567, "dur": 5.287,
    "args": {
      "External id": 261,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 66,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 66, "pid": 537, "tid": 537, "ts": 7429564822640.567,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822665.812, "dur": 3.800,
    "args": {
      "External id": 270,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 67,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 67, "pid": 537, "tid": 537, "ts": 7429564822665.812,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822715.824, "dur": 0.274,
    "args": {
      "External id": 280,      "cid": 370, "correlation": 68
    }
  },
  {
    "ph": "f", "id": 68, "pid": 537, "tid": 537, "ts": 7429564822715.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822718.808, "dur": 0.151,
    "args": {
      "External id": 280,      "cid": 370, "correlation": 69
    }
  },
  {
    "ph": "f", "id": 69, "pid": 537, "tid": 537, "ts": 7429564822718.808,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822721.193, "dur": 0.154,
    "args": {
      "External id": 280,      "cid": 370, "correlation": 70
    }
  },
  {
    "ph": "f", "id": 70, "pid": 537, "tid": 537, "ts": 7429564822721.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822723.111, "dur": 4.419,
    "args": {
      "External id": 280,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 71,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 71, "pid": 537, "tid": 537, "ts": 7429564822723.111,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822741.275, "dur": 4.826,
    "args": {
      "External id": 283,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 72,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 72, "pid": 537, "tid": 537, "ts": 7429564822741.275,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822767.517, "dur": 5.715,
    "args": {
      "External id": 285,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 73,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 73, "pid": 537, "tid": 537, "ts": 7429564822767.517,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822785.329, "dur": 5.662,
    "args": {
      "External id": 288,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 74,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 74, "pid": 537, "tid": 537, "ts": 7429564822785.329,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822799.379, "dur": 3.968,
    "args": {
      "External id": 289,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 75,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 75, "pid": 537, "tid": 537, "ts": 7429564822799.379,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822810.541, "dur": 4.247,
    "args": {
      "External id": 290,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 76,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 76, "pid": 537, "tid": 537, "ts": 7429564822810.541,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822822.277, "dur": 4.982,
    "args": {
      "External id": 291,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 77,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 77, "pid": 537, "tid": 537, "ts": 7429564822822.277,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822835.787, "dur": 3.996,
    "args": {
      "External id": 293,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 78,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 78, "pid": 537, "tid": 537, "ts": 7429564822835.787,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822881.436, "dur": 0.246,
    "args": {
      "External id": 301,      "cid": 370, "correlation": 79
    }
  },
  {
    "ph": "f", "id": 79, "pid": 537, "tid": 537, "ts": 7429564822881.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822884.599, "dur": 0.170,
    "args": {
      "External id": 301,      "cid": 370, "correlation": 80
    }
  },
  {
    "ph": "f", "id": 80, "pid": 537, "tid": 537, "ts": 7429564822884.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822887.161, "dur": 0.157,
    "args": {
      "External id": 301,      "cid": 370, "correlation": 81
    }
  },
  {
    "ph": "f", "id": 81, "pid": 537, "tid": 537, "ts": 7429564822887.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822894.523, "dur": 5.523,
    "args": {
      "External id": 301,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 82,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 82, "pid": 537, "tid": 537, "ts": 7429564822894.523,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822900.432, "dur": 4.725,
    "args": {
      "External id": 301,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 83,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 83, "pid": 537, "tid": 537, "ts": 7429564822900.432,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822930.858, "dur": 5.853,
    "args": {
      "External id": 304,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 84,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 84, "pid": 537, "tid": 537, "ts": 7429564822930.858,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822964.144, "dur": 0.162,
    "args": {
      "External id": 312,      "cid": 370, "correlation": 85
    }
  },
  {
    "ph": "f", "id": 85, "pid": 537, "tid": 537, "ts": 7429564822964.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822966.537, "dur": 0.156,
    "args": {
      "External id": 312,      "cid": 370, "correlation": 86
    }
  },
  {
    "ph": "f", "id": 86, "pid": 537, "tid": 537, "ts": 7429564822966.537,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564822968.455, "dur": 0.140,
    "args": {
      "External id": 312,      "cid": 370, "correlation": 87
    }
  },
  {
    "ph": "f", "id": 87, "pid": 537, "tid": 537, "ts": 7429564822968.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822972.409, "dur": 5.082,
    "args": {
      "External id": 312,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 88,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 88, "pid": 537, "tid": 537, "ts": 7429564822972.409,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822977.810, "dur": 3.745,
    "args": {
      "External id": 312,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 89,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 89, "pid": 537, "tid": 537, "ts": 7429564822977.810,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564822991.445, "dur": 4.566,
    "args": {
      "External id": 315,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 90,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 90, "pid": 537, "tid": 537, "ts": 7429564822991.445,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823022.816, "dur": 0.159,
    "args": {
      "External id": 323,      "cid": 370, "correlation": 91
    }
  },
  {
    "ph": "f", "id": 91, "pid": 537, "tid": 537, "ts": 7429564823022.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823025.373, "dur": 0.150,
    "args": {
      "External id": 323,      "cid": 370, "correlation": 92
    }
  },
  {
    "ph": "f", "id": 92, "pid": 537, "tid": 537, "ts": 7429564823025.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823027.970, "dur": 0.165,
    "args": {
      "External id": 323,      "cid": 370, "correlation": 93
    }
  },
  {
    "ph": "f", "id": 93, "pid": 537, "tid": 537, "ts": 7429564823027.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823031.530, "dur": 6.395,
    "args": {
      "External id": 323,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 94,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 94, "pid": 537, "tid": 537, "ts": 7429564823031.530,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823038.238, "dur": 8.967,
    "args": {
      "External id": 323,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 95,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 95, "pid": 537, "tid": 537, "ts": 7429564823038.238,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823057.306, "dur": 4.333,
    "args": {
      "External id": 326,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 96,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 96, "pid": 537, "tid": 537, "ts": 7429564823057.306,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823089.517, "dur": 4.596,
    "args": {
      "External id": 328,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 97,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 97, "pid": 537, "tid": 537, "ts": 7429564823089.517,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823102.727, "dur": 5.781,
    "args": {
      "External id": 331,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 98,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 98, "pid": 537, "tid": 537, "ts": 7429564823102.727,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823116.194, "dur": 4.099,
    "args": {
      "External id": 332,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 99,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 99, "pid": 537, "tid": 537, "ts": 7429564823116.194,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823125.609, "dur": 4.174,
    "args": {
      "External id": 333,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 100,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 100, "pid": 537, "tid": 537, "ts": 7429564823125.609,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823135.928, "dur": 6.006,
    "args": {
      "External id": 334,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 101,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 101, "pid": 537, "tid": 537, "ts": 7429564823135.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823149.199, "dur": 4.765,
    "args": {
      "External id": 336,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 102,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 102, "pid": 537, "tid": 537, "ts": 7429564823149.199,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823192.874, "dur": 0.324,
    "args": {
      "External id": 344,      "cid": 370, "correlation": 103
    }
  },
  {
    "ph": "f", "id": 103, "pid": 537, "tid": 537, "ts": 7429564823192.874,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823195.943, "dur": 0.159,
    "args": {
      "External id": 344,      "cid": 370, "correlation": 104
    }
  },
  {
    "ph": "f", "id": 104, "pid": 537, "tid": 537, "ts": 7429564823195.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823198.309, "dur": 0.169,
    "args": {
      "External id": 344,      "cid": 370, "correlation": 105
    }
  },
  {
    "ph": "f", "id": 105, "pid": 537, "tid": 537, "ts": 7429564823198.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823200.798, "dur": 5.580,
    "args": {
      "External id": 344,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 106,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 106, "pid": 537, "tid": 537, "ts": 7429564823200.798,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823243.731, "dur": 0.193,
    "args": {
      "External id": 357,      "cid": 370, "correlation": 107
    }
  },
  {
    "ph": "f", "id": 107, "pid": 537, "tid": 537, "ts": 7429564823243.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823246.202, "dur": 0.163,
    "args": {
      "External id": 357,      "cid": 370, "correlation": 108
    }
  },
  {
    "ph": "f", "id": 108, "pid": 537, "tid": 537, "ts": 7429564823246.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823248.140, "dur": 0.140,
    "args": {
      "External id": 357,      "cid": 370, "correlation": 109
    }
  },
  {
    "ph": "f", "id": 109, "pid": 537, "tid": 537, "ts": 7429564823248.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823252.429, "dur": 6.519,
    "args": {
      "External id": 357,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 110,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 110, "pid": 537, "tid": 537, "ts": 7429564823252.429,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823259.368, "dur": 4.437,
    "args": {
      "External id": 357,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 111,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 111, "pid": 537, "tid": 537, "ts": 7429564823259.368,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823298.627, "dur": 0.192,
    "args": {
      "External id": 370,      "cid": 370, "correlation": 112
    }
  },
  {
    "ph": "f", "id": 112, "pid": 537, "tid": 537, "ts": 7429564823298.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823300.887, "dur": 0.148,
    "args": {
      "External id": 370,      "cid": 370, "correlation": 113
    }
  },
  {
    "ph": "f", "id": 113, "pid": 537, "tid": 537, "ts": 7429564823300.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823303.150, "dur": 0.143,
    "args": {
      "External id": 370,      "cid": 370, "correlation": 114
    }
  },
  {
    "ph": "f", "id": 114, "pid": 537, "tid": 537, "ts": 7429564823303.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823306.673, "dur": 4.620,
    "args": {
      "External id": 370,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 115,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 115, "pid": 537, "tid": 537, "ts": 7429564823306.673,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823311.617, "dur": 4.993,
    "args": {
      "External id": 370,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 116,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 116, "pid": 537, "tid": 537, "ts": 7429564823311.617,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823339.806, "dur": 4.727,
    "args": {
      "External id": 380,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 117,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 117, "pid": 537, "tid": 537, "ts": 7429564823339.806,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823364.663, "dur": 4.709,
    "args": {
      "External id": 385,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 118,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 118, "pid": 537, "tid": 537, "ts": 7429564823364.663,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823380.927, "dur": 6.347,
    "args": {
      "External id": 386,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 119,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 119, "pid": 537, "tid": 537, "ts": 7429564823380.927,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823393.914, "dur": 5.638,
    "args": {
      "External id": 387,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 120,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 120, "pid": 537, "tid": 537, "ts": 7429564823393.914,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823405.611, "dur": 4.438,
    "args": {
      "External id": 388,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 121,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 121, "pid": 537, "tid": 537, "ts": 7429564823405.611,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823415.554, "dur": 5.381,
    "args": {
      "External id": 389,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 122,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 122, "pid": 537, "tid": 537, "ts": 7429564823415.554,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823433.881, "dur": 4.599,
    "args": {
      "External id": 394,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 123,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 123, "pid": 537, "tid": 537, "ts": 7429564823433.881,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823444.480, "dur": 6.303,
    "args": {
      "External id": 395,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 124,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 124, "pid": 537, "tid": 537, "ts": 7429564823444.480,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823457.122, "dur": 4.112,
    "args": {
      "External id": 396,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 125,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 125, "pid": 537, "tid": 537, "ts": 7429564823457.122,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823465.430, "dur": 4.500,
    "args": {
      "External id": 397,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 126,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 126, "pid": 537, "tid": 537, "ts": 7429564823465.430,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823525.156, "dur": 4.851,
    "args": {
      "External id": 414,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 127,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 127, "pid": 537, "tid": 537, "ts": 7429564823525.156,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823558.413, "dur": 5.914,
    "args": {
      "External id": 432,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 128,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 128, "pid": 537, "tid": 537, "ts": 7429564823558.413,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823597.148, "dur": 0.314,
    "args": {
      "External id": 445,      "cid": 370, "correlation": 129
    }
  },
  {
    "ph": "f", "id": 129, "pid": 537, "tid": 537, "ts": 7429564823597.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823602.295, "dur": 0.167,
    "args": {
      "External id": 445,      "cid": 370, "correlation": 130
    }
  },
  {
    "ph": "f", "id": 130, "pid": 537, "tid": 537, "ts": 7429564823602.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823604.890, "dur": 0.152,
    "args": {
      "External id": 445,      "cid": 370, "correlation": 131
    }
  },
  {
    "ph": "f", "id": 131, "pid": 537, "tid": 537, "ts": 7429564823604.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823607.397, "dur": 6.201,
    "args": {
      "External id": 445,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 132,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 132, "pid": 537, "tid": 537, "ts": 7429564823607.397,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823624.796, "dur": 5.594,
    "args": {
      "External id": 448,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 133,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 133, "pid": 537, "tid": 537, "ts": 7429564823624.796,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823644.791, "dur": 4.260,
    "args": {
      "External id": 455,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 134,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 134, "pid": 537, "tid": 537, "ts": 7429564823644.791,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823660.271, "dur": 5.870,
    "args": {
      "External id": 458,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 135,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 135, "pid": 537, "tid": 537, "ts": 7429564823660.271,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823705.667, "dur": 0.259,
    "args": {
      "External id": 470,      "cid": 370, "correlation": 136
    }
  },
  {
    "ph": "f", "id": 136, "pid": 537, "tid": 537, "ts": 7429564823705.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823709.203, "dur": 0.151,
    "args": {
      "External id": 470,      "cid": 370, "correlation": 137
    }
  },
  {
    "ph": "f", "id": 137, "pid": 537, "tid": 537, "ts": 7429564823709.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823711.928, "dur": 0.149,
    "args": {
      "External id": 470,      "cid": 370, "correlation": 138
    }
  },
  {
    "ph": "f", "id": 138, "pid": 537, "tid": 537, "ts": 7429564823711.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823714.161, "dur": 5.350,
    "args": {
      "External id": 470,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 139,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 139, "pid": 537, "tid": 537, "ts": 7429564823714.161,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823736.321, "dur": 5.935,
    "args": {
      "External id": 479,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 140,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 140, "pid": 537, "tid": 537, "ts": 7429564823736.321,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823783.087, "dur": 0.275,
    "args": {
      "External id": 489,      "cid": 370, "correlation": 141
    }
  },
  {
    "ph": "f", "id": 141, "pid": 537, "tid": 537, "ts": 7429564823783.087,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823786.371, "dur": 0.157,
    "args": {
      "External id": 489,      "cid": 370, "correlation": 142
    }
  },
  {
    "ph": "f", "id": 142, "pid": 537, "tid": 537, "ts": 7429564823786.371,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564823788.530, "dur": 0.318,
    "args": {
      "External id": 489,      "cid": 370, "correlation": 143
    }
  },
  {
    "ph": "f", "id": 143, "pid": 537, "tid": 537, "ts": 7429564823788.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823790.591, "dur": 6.309,
    "args": {
      "External id": 489,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 144,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 144, "pid": 537, "tid": 537, "ts": 7429564823790.591,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564823812.361, "dur": 0.443,
    "args": {
      "External id": 492,      "cid": 290, "correlation": 145
    }
  },
  {
    "ph": "f", "id": 145, "pid": 537, "tid": 537, "ts": 7429564823812.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564823813.616, "dur": 72.140,
    "args": {
      "External id": 492,            "bytes": 2097152, "cid": 108, "correlation": 146, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 146, "pid": 537, "tid": 537, "ts": 7429564823813.616,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823888.980, "dur": 6.926,
    "args": {
      "External id": 492,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 147,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 147, "pid": 537, "tid": 537, "ts": 7429564823888.980,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823920.114, "dur": 6.118,
    "args": {
      "External id": 494,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 148,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 148, "pid": 537, "tid": 537, "ts": 7429564823920.114,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823937.721, "dur": 6.915,
    "args": {
      "External id": 497,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 149,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 149, "pid": 537, "tid": 537, "ts": 7429564823937.721,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823953.977, "dur": 24.603,
    "args": {
      "External id": 498,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 150,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 150, "pid": 537, "tid": 537, "ts": 7429564823953.977,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564823986.408, "dur": 5.903,
    "args": {
      "External id": 499,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 151,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 151, "pid": 537, "tid": 537, "ts": 7429564823986.408,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824000.401, "dur": 4.596,
    "args": {
      "External id": 500,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 152,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 152, "pid": 537, "tid": 537, "ts": 7429564824000.401,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824012.824, "dur": 3.480,
    "args": {
      "External id": 502,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 153,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 153, "pid": 537, "tid": 537, "ts": 7429564824012.824,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824057.306, "dur": 0.282,
    "args": {
      "External id": 510,      "cid": 370, "correlation": 154
    }
  },
  {
    "ph": "f", "id": 154, "pid": 537, "tid": 537, "ts": 7429564824057.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824061.447, "dur": 0.151,
    "args": {
      "External id": 510,      "cid": 370, "correlation": 155
    }
  },
  {
    "ph": "f", "id": 155, "pid": 537, "tid": 537, "ts": 7429564824061.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824064.086, "dur": 0.179,
    "args": {
      "External id": 510,      "cid": 370, "correlation": 156
    }
  },
  {
    "ph": "f", "id": 156, "pid": 537, "tid": 537, "ts": 7429564824064.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824073.240, "dur": 5.648,
    "args": {
      "External id": 510,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 157,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 157, "pid": 537, "tid": 537, "ts": 7429564824073.240,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824079.403, "dur": 4.286,
    "args": {
      "External id": 510,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 158,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 158, "pid": 537, "tid": 537, "ts": 7429564824079.403,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824150.865, "dur": 5.691,
    "args": {
      "External id": 513,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 159,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 159, "pid": 537, "tid": 537, "ts": 7429564824150.865,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564824177.655, "dur": 0.440,
    "args": {
      "External id": 521,      "cid": 290, "correlation": 160
    }
  },
  {
    "ph": "f", "id": 160, "pid": 537, "tid": 537, "ts": 7429564824177.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564824178.604, "dur": 47.775,
    "args": {
      "External id": 521,            "bytes": 2097152, "cid": 108, "correlation": 161, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 161, "pid": 537, "tid": 537, "ts": 7429564824178.604,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824239.322, "dur": 0.242,
    "args": {
      "External id": 521,      "cid": 370, "correlation": 162
    }
  },
  {
    "ph": "f", "id": 162, "pid": 537, "tid": 537, "ts": 7429564824239.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824243.703, "dur": 0.152,
    "args": {
      "External id": 521,      "cid": 370, "correlation": 163
    }
  },
  {
    "ph": "f", "id": 163, "pid": 537, "tid": 537, "ts": 7429564824243.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824246.289, "dur": 0.154,
    "args": {
      "External id": 521,      "cid": 370, "correlation": 164
    }
  },
  {
    "ph": "f", "id": 164, "pid": 537, "tid": 537, "ts": 7429564824246.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824253.323, "dur": 4.262,
    "args": {
      "External id": 521,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 165,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 165, "pid": 537, "tid": 537, "ts": 7429564824253.323,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824257.951, "dur": 4.534,
    "args": {
      "External id": 521,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 166,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 166, "pid": 537, "tid": 537, "ts": 7429564824257.951,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824274.433, "dur": 4.537,
    "args": {
      "External id": 524,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 167,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 167, "pid": 537, "tid": 537, "ts": 7429564824274.433,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824306.802, "dur": 0.190,
    "args": {
      "External id": 532,      "cid": 370, "correlation": 168
    }
  },
  {
    "ph": "f", "id": 168, "pid": 537, "tid": 537, "ts": 7429564824306.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824310.202, "dur": 0.154,
    "args": {
      "External id": 532,      "cid": 370, "correlation": 169
    }
  },
  {
    "ph": "f", "id": 169, "pid": 537, "tid": 537, "ts": 7429564824310.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824312.452, "dur": 0.148,
    "args": {
      "External id": 532,      "cid": 370, "correlation": 170
    }
  },
  {
    "ph": "f", "id": 170, "pid": 537, "tid": 537, "ts": 7429564824312.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824316.180, "dur": 4.796,
    "args": {
      "External id": 532,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 171,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 171, "pid": 537, "tid": 537, "ts": 7429564824316.180,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824321.263, "dur": 3.875,
    "args": {
      "External id": 532,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 172,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 172, "pid": 537, "tid": 537, "ts": 7429564824321.263,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824335.245, "dur": 4.541,
    "args": {
      "External id": 535,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 173,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 173, "pid": 537, "tid": 537, "ts": 7429564824335.245,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824370.669, "dur": 3.776,
    "args": {
      "External id": 537,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 174,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 174, "pid": 537, "tid": 537, "ts": 7429564824370.669,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824384.780, "dur": 4.760,
    "args": {
      "External id": 540,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 175,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 175, "pid": 537, "tid": 537, "ts": 7429564824384.780,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824398.688, "dur": 3.528,
    "args": {
      "External id": 541,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 176,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 176, "pid": 537, "tid": 537, "ts": 7429564824398.688,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824407.397, "dur": 4.419,
    "args": {
      "External id": 542,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 177,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 177, "pid": 537, "tid": 537, "ts": 7429564824407.397,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824417.990, "dur": 4.463,
    "args": {
      "External id": 543,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 178,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 178, "pid": 537, "tid": 537, "ts": 7429564824417.990,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564824431.328, "dur": 0.150,
    "args": {
      "External id": 545,      "cid": 290, "correlation": 179
    }
  },
  {
    "ph": "f", "id": 179, "pid": 537, "tid": 537, "ts": 7429564824431.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564824431.775, "dur": 37.084,
    "args": {
      "External id": 545,            "bytes": 2097152, "cid": 108, "correlation": 180, "ptr": "0x100000002"
    }
  },
  {
    "ph": "s", "id": 180, "pid": 537, "tid": 537, "ts": 7429564824431.775,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824471.057, "dur": 5.967,
    "args": {
      "External id": 545,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 181,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 181, "pid": 537, "tid": 537, "ts": 7429564824471.057,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824526.785, "dur": 0.313,
    "args": {
      "External id": 553,      "cid": 370, "correlation": 182
    }
  },
  {
    "ph": "f", "id": 182, "pid": 537, "tid": 537, "ts": 7429564824526.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824531.435, "dur": 0.166,
    "args": {
      "External id": 553,      "cid": 370, "correlation": 183
    }
  },
  {
    "ph": "f", "id": 183, "pid": 537, "tid": 537, "ts": 7429564824531.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824534.004, "dur": 0.163,
    "args": {
      "External id": 553,      "cid": 370, "correlation": 184
    }
  },
  {
    "ph": "f", "id": 184, "pid": 537, "tid": 537, "ts": 7429564824534.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824536.696, "dur": 5.351,
    "args": {
      "External id": 553,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 185,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 185, "pid": 537, "tid": 537, "ts": 7429564824536.696,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824581.528, "dur": 0.254,
    "args": {
      "External id": 566,      "cid": 370, "correlation": 186
    }
  },
  {
    "ph": "f", "id": 186, "pid": 537, "tid": 537, "ts": 7429564824581.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824585.133, "dur": 0.161,
    "args": {
      "External id": 566,      "cid": 370, "correlation": 187
    }
  },
  {
    "ph": "f", "id": 187, "pid": 537, "tid": 537, "ts": 7429564824585.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824587.228, "dur": 0.167,
    "args": {
      "External id": 566,      "cid": 370, "correlation": 188
    }
  },
  {
    "ph": "f", "id": 188, "pid": 537, "tid": 537, "ts": 7429564824587.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824592.609, "dur": 5.380,
    "args": {
      "External id": 566,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 189,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 189, "pid": 537, "tid": 537, "ts": 7429564824592.609,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824598.510, "dur": 3.821,
    "args": {
      "External id": 566,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 190,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 190, "pid": 537, "tid": 537, "ts": 7429564824598.510,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824636.750, "dur": 0.219,
    "args": {
      "External id": 579,      "cid": 370, "correlation": 191
    }
  },
  {
    "ph": "f", "id": 191, "pid": 537, "tid": 537, "ts": 7429564824636.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824639.763, "dur": 0.141,
    "args": {
      "External id": 579,      "cid": 370, "correlation": 192
    }
  },
  {
    "ph": "f", "id": 192, "pid": 537, "tid": 537, "ts": 7429564824639.763,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824641.773, "dur": 0.152,
    "args": {
      "External id": 579,      "cid": 370, "correlation": 193
    }
  },
  {
    "ph": "f", "id": 193, "pid": 537, "tid": 537, "ts": 7429564824641.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824645.238, "dur": 4.290,
    "args": {
      "External id": 579,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 194,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 194, "pid": 537, "tid": 537, "ts": 7429564824645.238,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824649.932, "dur": 3.974,
    "args": {
      "External id": 579,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 195,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 195, "pid": 537, "tid": 537, "ts": 7429564824649.932,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824678.704, "dur": 4.217,
    "args": {
      "External id": 589,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 196,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 196, "pid": 537, "tid": 537, "ts": 7429564824678.704,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824709.474, "dur": 4.947,
    "args": {
      "External id": 594,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 197,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 197, "pid": 537, "tid": 537, "ts": 7429564824709.474,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824730.912, "dur": 7.429,
    "args": {
      "External id": 595,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 198,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 198, "pid": 537, "tid": 537, "ts": 7429564824730.912,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824747.040, "dur": 3.788,
    "args": {
      "External id": 596,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 199,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 199, "pid": 537, "tid": 537, "ts": 7429564824747.040,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824756.454, "dur": 4.595,
    "args": {
      "External id": 597,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 200,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 200, "pid": 537, "tid": 537, "ts": 7429564824756.454,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824766.718, "dur": 4.156,
    "args": {
      "External id": 598,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 201,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 201, "pid": 537, "tid": 537, "ts": 7429564824766.718,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824785.030, "dur": 4.171,
    "args": {
      "External id": 603,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 202,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 202, "pid": 537, "tid": 537, "ts": 7429564824785.030,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824796.833, "dur": 5.717,
    "args": {
      "External id": 604,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 203,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 203, "pid": 537, "tid": 537, "ts": 7429564824796.833,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824808.160, "dur": 3.206,
    "args": {
      "External id": 605,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 204,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 204, "pid": 537, "tid": 537, "ts": 7429564824808.160,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824815.900, "dur": 3.311,
    "args": {
      "External id": 606,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 205,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 205, "pid": 537, "tid": 537, "ts": 7429564824815.900,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824883.383, "dur": 4.955,
    "args": {
      "External id": 623,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 206,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 206, "pid": 537, "tid": 537, "ts": 7429564824883.383,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824920.713, "dur": 3.727,
    "args": {
      "External id": 641,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 207,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 207, "pid": 537, "tid": 537, "ts": 7429564824920.713,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824959.035, "dur": 0.317,
    "args": {
      "External id": 654,      "cid": 370, "correlation": 208
    }
  },
  {
    "ph": "f", "id": 208, "pid": 537, "tid": 537, "ts": 7429564824959.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824963.162, "dur": 0.162,
    "args": {
      "External id": 654,      "cid": 370, "correlation": 209
    }
  },
  {
    "ph": "f", "id": 209, "pid": 537, "tid": 537, "ts": 7429564824963.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564824965.528, "dur": 0.149,
    "args": {
      "External id": 654,      "cid": 370, "correlation": 210
    }
  },
  {
    "ph": "f", "id": 210, "pid": 537, "tid": 537, "ts": 7429564824965.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824968.172, "dur": 5.338,
    "args": {
      "External id": 654,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 211,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 211, "pid": 537, "tid": 537, "ts": 7429564824968.172,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564824985.605, "dur": 4.306,
    "args": {
      "External id": 657,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 212,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 212, "pid": 537, "tid": 537, "ts": 7429564824985.605,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825004.415, "dur": 3.485,
    "args": {
      "External id": 664,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 213,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 213, "pid": 537, "tid": 537, "ts": 7429564825004.415,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825020.589, "dur": 5.701,
    "args": {
      "External id": 667,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 214,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 214, "pid": 537, "tid": 537, "ts": 7429564825020.589,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825073.154, "dur": 0.192,
    "args": {
      "External id": 679,      "cid": 370, "correlation": 215
    }
  },
  {
    "ph": "f", "id": 215, "pid": 537, "tid": 537, "ts": 7429564825073.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825076.347, "dur": 0.144,
    "args": {
      "External id": 679,      "cid": 370, "correlation": 216
    }
  },
  {
    "ph": "f", "id": 216, "pid": 537, "tid": 537, "ts": 7429564825076.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825079.018, "dur": 0.158,
    "args": {
      "External id": 679,      "cid": 370, "correlation": 217
    }
  },
  {
    "ph": "f", "id": 217, "pid": 537, "tid": 537, "ts": 7429564825079.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825081.121, "dur": 4.674,
    "args": {
      "External id": 679,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 218,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 218, "pid": 537, "tid": 537, "ts": 7429564825081.121,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825101.863, "dur": 4.404,
    "args": {
      "External id": 688,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 219,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 219, "pid": 537, "tid": 537, "ts": 7429564825101.863,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825147.114, "dur": 0.182,
    "args": {
      "External id": 698,      "cid": 370, "correlation": 220
    }
  },
  {
    "ph": "f", "id": 220, "pid": 537, "tid": 537, "ts": 7429564825147.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825150.124, "dur": 0.153,
    "args": {
      "External id": 698,      "cid": 370, "correlation": 221
    }
  },
  {
    "ph": "f", "id": 221, "pid": 537, "tid": 537, "ts": 7429564825150.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825152.369, "dur": 0.188,
    "args": {
      "External id": 698,      "cid": 370, "correlation": 222
    }
  },
  {
    "ph": "f", "id": 222, "pid": 537, "tid": 537, "ts": 7429564825152.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825154.266, "dur": 4.488,
    "args": {
      "External id": 698,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 223,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 223, "pid": 537, "tid": 537, "ts": 7429564825154.266,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825170.459, "dur": 4.720,
    "args": {
      "External id": 701,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 224,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 224, "pid": 537, "tid": 537, "ts": 7429564825170.459,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825192.797, "dur": 4.691,
    "args": {
      "External id": 703,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 225,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 225, "pid": 537, "tid": 537, "ts": 7429564825192.797,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825207.774, "dur": 4.901,
    "args": {
      "External id": 706,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 226,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 226, "pid": 537, "tid": 537, "ts": 7429564825207.774,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825222.018, "dur": 3.550,
    "args": {
      "External id": 707,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 227,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 227, "pid": 537, "tid": 537, "ts": 7429564825222.018,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825231.681, "dur": 5.078,
    "args": {
      "External id": 708,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 228,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 228, "pid": 537, "tid": 537, "ts": 7429564825231.681,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825243.607, "dur": 4.862,
    "args": {
      "External id": 709,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 229,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 229, "pid": 537, "tid": 537, "ts": 7429564825243.607,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825256.237, "dur": 3.709,
    "args": {
      "External id": 711,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 230,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 230, "pid": 537, "tid": 537, "ts": 7429564825256.237,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564825287.684, "dur": 0.351,
    "args": {
      "External id": 719,      "cid": 290, "correlation": 231
    }
  },
  {
    "ph": "f", "id": 231, "pid": 537, "tid": 537, "ts": 7429564825287.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564825288.454, "dur": 56.600,
    "args": {
      "External id": 719,            "bytes": 2097152, "cid": 108, "correlation": 232, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 232, "pid": 537, "tid": 537, "ts": 7429564825288.454,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825360.618, "dur": 0.345,
    "args": {
      "External id": 719,      "cid": 370, "correlation": 233
    }
  },
  {
    "ph": "f", "id": 233, "pid": 537, "tid": 537, "ts": 7429564825360.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825365.541, "dur": 0.153,
    "args": {
      "External id": 719,      "cid": 370, "correlation": 234
    }
  },
  {
    "ph": "f", "id": 234, "pid": 537, "tid": 537, "ts": 7429564825365.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825368.338, "dur": 0.150,
    "args": {
      "External id": 719,      "cid": 370, "correlation": 235
    }
  },
  {
    "ph": "f", "id": 235, "pid": 537, "tid": 537, "ts": 7429564825368.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825377.140, "dur": 6.071,
    "args": {
      "External id": 719,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 236,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 236, "pid": 537, "tid": 537, "ts": 7429564825377.140,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825383.601, "dur": 4.008,
    "args": {
      "External id": 719,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 237,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 237, "pid": 537, "tid": 537, "ts": 7429564825383.601,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825409.745, "dur": 4.694,
    "args": {
      "External id": 722,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 238,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 238, "pid": 537, "tid": 537, "ts": 7429564825409.745,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564825435.937, "dur": 0.192,
    "args": {
      "External id": 730,      "cid": 290, "correlation": 239
    }
  },
  {
    "ph": "f", "id": 239, "pid": 537, "tid": 537, "ts": 7429564825435.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564825436.476, "dur": 31.506,
    "args": {
      "External id": 730,            "bytes": 2097152, "cid": 108, "correlation": 240, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 240, "pid": 537, "tid": 537, "ts": 7429564825436.476,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825481.308, "dur": 0.225,
    "args": {
      "External id": 730,      "cid": 370, "correlation": 241
    }
  },
  {
    "ph": "f", "id": 241, "pid": 537, "tid": 537, "ts": 7429564825481.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825484.611, "dur": 0.156,
    "args": {
      "External id": 730,      "cid": 370, "correlation": 242
    }
  },
  {
    "ph": "f", "id": 242, "pid": 537, "tid": 537, "ts": 7429564825484.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825487.237, "dur": 0.157,
    "args": {
      "External id": 730,      "cid": 370, "correlation": 243
    }
  },
  {
    "ph": "f", "id": 243, "pid": 537, "tid": 537, "ts": 7429564825487.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825492.165, "dur": 5.735,
    "args": {
      "External id": 730,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 244,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 244, "pid": 537, "tid": 537, "ts": 7429564825492.165,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825498.225, "dur": 5.216,
    "args": {
      "External id": 730,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 245,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 245, "pid": 537, "tid": 537, "ts": 7429564825498.225,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825514.638, "dur": 4.206,
    "args": {
      "External id": 733,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 246,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 246, "pid": 537, "tid": 537, "ts": 7429564825514.638,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825545.687, "dur": 0.156,
    "args": {
      "External id": 741,      "cid": 370, "correlation": 247
    }
  },
  {
    "ph": "f", "id": 247, "pid": 537, "tid": 537, "ts": 7429564825545.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825549.376, "dur": 0.149,
    "args": {
      "External id": 741,      "cid": 370, "correlation": 248
    }
  },
  {
    "ph": "f", "id": 248, "pid": 537, "tid": 537, "ts": 7429564825549.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825551.449, "dur": 0.148,
    "args": {
      "External id": 741,      "cid": 370, "correlation": 249
    }
  },
  {
    "ph": "f", "id": 249, "pid": 537, "tid": 537, "ts": 7429564825551.449,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825555.497, "dur": 4.751,
    "args": {
      "External id": 741,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 250,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 250, "pid": 537, "tid": 537, "ts": 7429564825555.497,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825560.659, "dur": 3.677,
    "args": {
      "External id": 741,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 251,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 251, "pid": 537, "tid": 537, "ts": 7429564825560.659,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825575.176, "dur": 5.070,
    "args": {
      "External id": 744,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 252,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 252, "pid": 537, "tid": 537, "ts": 7429564825575.176,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825610.091, "dur": 4.720,
    "args": {
      "External id": 746,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 253,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 253, "pid": 537, "tid": 537, "ts": 7429564825610.091,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825625.030, "dur": 5.620,
    "args": {
      "External id": 749,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 254,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 254, "pid": 537, "tid": 537, "ts": 7429564825625.030,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825639.341, "dur": 3.937,
    "args": {
      "External id": 750,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 255,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 255, "pid": 537, "tid": 537, "ts": 7429564825639.341,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825649.372, "dur": 5.110,
    "args": {
      "External id": 751,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 256,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 256, "pid": 537, "tid": 537, "ts": 7429564825649.372,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825661.042, "dur": 3.768,
    "args": {
      "External id": 752,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 257,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 257, "pid": 537, "tid": 537, "ts": 7429564825661.042,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825675.442, "dur": 3.631,
    "args": {
      "External id": 754,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 258,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 258, "pid": 537, "tid": 537, "ts": 7429564825675.442,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825720.160, "dur": 0.190,
    "args": {
      "External id": 762,      "cid": 370, "correlation": 259
    }
  },
  {
    "ph": "f", "id": 259, "pid": 537, "tid": 537, "ts": 7429564825720.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825724.361, "dur": 0.150,
    "args": {
      "External id": 762,      "cid": 370, "correlation": 260
    }
  },
  {
    "ph": "f", "id": 260, "pid": 537, "tid": 537, "ts": 7429564825724.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825726.923, "dur": 0.155,
    "args": {
      "External id": 762,      "cid": 370, "correlation": 261
    }
  },
  {
    "ph": "f", "id": 261, "pid": 537, "tid": 537, "ts": 7429564825726.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825729.152, "dur": 4.841,
    "args": {
      "External id": 762,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 262,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 262, "pid": 537, "tid": 537, "ts": 7429564825729.152,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825772.282, "dur": 0.195,
    "args": {
      "External id": 775,      "cid": 370, "correlation": 263
    }
  },
  {
    "ph": "f", "id": 263, "pid": 537, "tid": 537, "ts": 7429564825772.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825775.458, "dur": 0.144,
    "args": {
      "External id": 775,      "cid": 370, "correlation": 264
    }
  },
  {
    "ph": "f", "id": 264, "pid": 537, "tid": 537, "ts": 7429564825775.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825777.364, "dur": 0.156,
    "args": {
      "External id": 775,      "cid": 370, "correlation": 265
    }
  },
  {
    "ph": "f", "id": 265, "pid": 537, "tid": 537, "ts": 7429564825777.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825781.334, "dur": 4.945,
    "args": {
      "External id": 775,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 266,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 266, "pid": 537, "tid": 537, "ts": 7429564825781.334,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825786.680, "dur": 4.323,
    "args": {
      "External id": 775,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 267,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 267, "pid": 537, "tid": 537, "ts": 7429564825786.680,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825823.457, "dur": 0.166,
    "args": {
      "External id": 788,      "cid": 370, "correlation": 268
    }
  },
  {
    "ph": "f", "id": 268, "pid": 537, "tid": 537, "ts": 7429564825823.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825826.012, "dur": 0.147,
    "args": {
      "External id": 788,      "cid": 370, "correlation": 269
    }
  },
  {
    "ph": "f", "id": 269, "pid": 537, "tid": 537, "ts": 7429564825826.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564825827.833, "dur": 0.161,
    "args": {
      "External id": 788,      "cid": 370, "correlation": 270
    }
  },
  {
    "ph": "f", "id": 270, "pid": 537, "tid": 537, "ts": 7429564825827.833,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825830.991, "dur": 4.631,
    "args": {
      "External id": 788,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 271,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 271, "pid": 537, "tid": 537, "ts": 7429564825830.991,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825835.913, "dur": 3.339,
    "args": {
      "External id": 788,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 272,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 272, "pid": 537, "tid": 537, "ts": 7429564825835.913,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825862.314, "dur": 3.681,
    "args": {
      "External id": 798,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 273,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 273, "pid": 537, "tid": 537, "ts": 7429564825862.314,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825891.365, "dur": 4.819,
    "args": {
      "External id": 803,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 274,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 274, "pid": 537, "tid": 537, "ts": 7429564825891.365,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564825907.345, "dur": 0.328,
    "args": {
      "External id": 804,      "cid": 290, "correlation": 275
    }
  },
  {
    "ph": "f", "id": 275, "pid": 537, "tid": 537, "ts": 7429564825907.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564825908.182, "dur": 38.732,
    "args": {
      "External id": 804,            "bytes": 2097152, "cid": 108, "correlation": 276, "ptr": "0x5569ba9ecb70"
    }
  },
  {
    "ph": "s", "id": 276, "pid": 537, "tid": 537, "ts": 7429564825908.182,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825952.193, "dur": 7.394,
    "args": {
      "External id": 804,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 277,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 277, "pid": 537, "tid": 537, "ts": 7429564825952.193,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825969.683, "dur": 3.961,
    "args": {
      "External id": 805,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 278,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 278, "pid": 537, "tid": 537, "ts": 7429564825969.683,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825980.022, "dur": 4.332,
    "args": {
      "External id": 806,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 279,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 279, "pid": 537, "tid": 537, "ts": 7429564825980.022,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564825991.374, "dur": 3.310,
    "args": {
      "External id": 807,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 280,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 280, "pid": 537, "tid": 537, "ts": 7429564825991.374,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826009.306, "dur": 4.721,
    "args": {
      "External id": 812,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 281,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 281, "pid": 537, "tid": 537, "ts": 7429564826009.306,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826021.463, "dur": 5.708,
    "args": {
      "External id": 813,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 282,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 282, "pid": 537, "tid": 537, "ts": 7429564826021.463,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826032.935, "dur": 4.308,
    "args": {
      "External id": 814,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 283,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 283, "pid": 537, "tid": 537, "ts": 7429564826032.935,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826042.892, "dur": 4.132,
    "args": {
      "External id": 815,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 284,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 284, "pid": 537, "tid": 537, "ts": 7429564826042.892,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826134.166, "dur": 5.139,
    "args": {
      "External id": 832,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 285,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 285, "pid": 537, "tid": 537, "ts": 7429564826134.166,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826172.440, "dur": 4.042,
    "args": {
      "External id": 850,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 286,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 286, "pid": 537, "tid": 537, "ts": 7429564826172.440,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826217.532, "dur": 0.398,
    "args": {
      "External id": 863,      "cid": 370, "correlation": 287
    }
  },
  {
    "ph": "f", "id": 287, "pid": 537, "tid": 537, "ts": 7429564826217.532,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826223.441, "dur": 0.167,
    "args": {
      "External id": 863,      "cid": 370, "correlation": 288
    }
  },
  {
    "ph": "f", "id": 288, "pid": 537, "tid": 537, "ts": 7429564826223.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826226.311, "dur": 0.168,
    "args": {
      "External id": 863,      "cid": 370, "correlation": 289
    }
  },
  {
    "ph": "f", "id": 289, "pid": 537, "tid": 537, "ts": 7429564826226.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826230.374, "dur": 5.061,
    "args": {
      "External id": 863,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 290,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 290, "pid": 537, "tid": 537, "ts": 7429564826230.374,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826249.911, "dur": 4.973,
    "args": {
      "External id": 866,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 291,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 291, "pid": 537, "tid": 537, "ts": 7429564826249.911,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826270.687, "dur": 4.732,
    "args": {
      "External id": 873,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 292,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 292, "pid": 537, "tid": 537, "ts": 7429564826270.687,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826288.186, "dur": 5.870,
    "args": {
      "External id": 876,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 293,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 293, "pid": 537, "tid": 537, "ts": 7429564826288.186,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826338.953, "dur": 0.193,
    "args": {
      "External id": 888,      "cid": 370, "correlation": 294
    }
  },
  {
    "ph": "f", "id": 294, "pid": 537, "tid": 537, "ts": 7429564826338.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826342.892, "dur": 0.148,
    "args": {
      "External id": 888,      "cid": 370, "correlation": 295
    }
  },
  {
    "ph": "f", "id": 295, "pid": 537, "tid": 537, "ts": 7429564826342.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826345.995, "dur": 0.145,
    "args": {
      "External id": 888,      "cid": 370, "correlation": 296
    }
  },
  {
    "ph": "f", "id": 296, "pid": 537, "tid": 537, "ts": 7429564826345.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826348.243, "dur": 6.128,
    "args": {
      "External id": 888,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 297,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 297, "pid": 537, "tid": 537, "ts": 7429564826348.243,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826371.736, "dur": 4.069,
    "args": {
      "External id": 897,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 298,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 298, "pid": 537, "tid": 537, "ts": 7429564826371.736,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826421.158, "dur": 0.294,
    "args": {
      "External id": 907,      "cid": 370, "correlation": 299
    }
  },
  {
    "ph": "f", "id": 299, "pid": 537, "tid": 537, "ts": 7429564826421.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826424.693, "dur": 0.214,
    "args": {
      "External id": 907,      "cid": 370, "correlation": 300
    }
  },
  {
    "ph": "f", "id": 300, "pid": 537, "tid": 537, "ts": 7429564826424.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826426.872, "dur": 0.295,
    "args": {
      "External id": 907,      "cid": 370, "correlation": 301
    }
  },
  {
    "ph": "f", "id": 301, "pid": 537, "tid": 537, "ts": 7429564826426.872,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826428.868, "dur": 4.785,
    "args": {
      "External id": 907,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 302,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 302, "pid": 537, "tid": 537, "ts": 7429564826428.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826446.555, "dur": 4.652,
    "args": {
      "External id": 910,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 303,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 303, "pid": 537, "tid": 537, "ts": 7429564826446.555,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826470.246, "dur": 5.463,
    "args": {
      "External id": 912,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 304,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 304, "pid": 537, "tid": 537, "ts": 7429564826470.246,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826486.627, "dur": 4.977,
    "args": {
      "External id": 915,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 305,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 305, "pid": 537, "tid": 537, "ts": 7429564826486.627,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826501.099, "dur": 4.684,
    "args": {
      "External id": 916,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 306,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 306, "pid": 537, "tid": 537, "ts": 7429564826501.099,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826512.609, "dur": 4.139,
    "args": {
      "External id": 917,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 307,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 307, "pid": 537, "tid": 537, "ts": 7429564826512.609,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826523.324, "dur": 4.464,
    "args": {
      "External id": 918,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 308,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 308, "pid": 537, "tid": 537, "ts": 7429564826523.324,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826536.645, "dur": 4.389,
    "args": {
      "External id": 920,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 309,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 309, "pid": 537, "tid": 537, "ts": 7429564826536.645,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826575.180, "dur": 0.195,
    "args": {
      "External id": 928,      "cid": 370, "correlation": 310
    }
  },
  {
    "ph": "f", "id": 310, "pid": 537, "tid": 537, "ts": 7429564826575.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826578.588, "dur": 0.141,
    "args": {
      "External id": 928,      "cid": 370, "correlation": 311
    }
  },
  {
    "ph": "f", "id": 311, "pid": 537, "tid": 537, "ts": 7429564826578.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826581.023, "dur": 0.150,
    "args": {
      "External id": 928,      "cid": 370, "correlation": 312
    }
  },
  {
    "ph": "f", "id": 312, "pid": 537, "tid": 537, "ts": 7429564826581.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826589.052, "dur": 5.692,
    "args": {
      "External id": 928,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 313,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 313, "pid": 537, "tid": 537, "ts": 7429564826589.052,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826595.298, "dur": 3.740,
    "args": {
      "External id": 928,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 314,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 314, "pid": 537, "tid": 537, "ts": 7429564826595.298,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564826615.957, "dur": 0.353,
    "args": {
      "External id": 931,      "cid": 290, "correlation": 315
    }
  },
  {
    "ph": "f", "id": 315, "pid": 537, "tid": 537, "ts": 7429564826615.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564826616.928, "dur": 52.306,
    "args": {
      "External id": 931,            "bytes": 2097152, "cid": 108, "correlation": 316, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 316, "pid": 537, "tid": 537, "ts": 7429564826616.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826672.598, "dur": 4.973,
    "args": {
      "External id": 931,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 317,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 317, "pid": 537, "tid": 537, "ts": 7429564826672.598,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826710.835, "dur": 0.269,
    "args": {
      "External id": 939,      "cid": 370, "correlation": 318
    }
  },
  {
    "ph": "f", "id": 318, "pid": 537, "tid": 537, "ts": 7429564826710.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826714.093, "dur": 0.162,
    "args": {
      "External id": 939,      "cid": 370, "correlation": 319
    }
  },
  {
    "ph": "f", "id": 319, "pid": 537, "tid": 537, "ts": 7429564826714.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826716.549, "dur": 0.144,
    "args": {
      "External id": 939,      "cid": 370, "correlation": 320
    }
  },
  {
    "ph": "f", "id": 320, "pid": 537, "tid": 537, "ts": 7429564826716.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826721.588, "dur": 4.944,
    "args": {
      "External id": 939,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 321,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 321, "pid": 537, "tid": 537, "ts": 7429564826721.588,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826727.032, "dur": 3.278,
    "args": {
      "External id": 939,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 322,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 322, "pid": 537, "tid": 537, "ts": 7429564826727.032,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564826739.059, "dur": 0.217,
    "args": {
      "External id": 942,      "cid": 290, "correlation": 323
    }
  },
  {
    "ph": "f", "id": 323, "pid": 537, "tid": 537, "ts": 7429564826739.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564826739.719, "dur": 27.377,
    "args": {
      "External id": 942,            "bytes": 2097152, "cid": 108, "correlation": 324, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 324, "pid": 537, "tid": 537, "ts": 7429564826739.719,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826769.494, "dur": 4.145,
    "args": {
      "External id": 942,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 325,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 325, "pid": 537, "tid": 537, "ts": 7429564826769.494,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826802.177, "dur": 0.191,
    "args": {
      "External id": 950,      "cid": 370, "correlation": 326
    }
  },
  {
    "ph": "f", "id": 326, "pid": 537, "tid": 537, "ts": 7429564826802.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826805.772, "dur": 0.153,
    "args": {
      "External id": 950,      "cid": 370, "correlation": 327
    }
  },
  {
    "ph": "f", "id": 327, "pid": 537, "tid": 537, "ts": 7429564826805.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826808.218, "dur": 0.145,
    "args": {
      "External id": 950,      "cid": 370, "correlation": 328
    }
  },
  {
    "ph": "f", "id": 328, "pid": 537, "tid": 537, "ts": 7429564826808.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826812.522, "dur": 5.664,
    "args": {
      "External id": 950,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 329,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 329, "pid": 537, "tid": 537, "ts": 7429564826812.522,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826818.653, "dur": 3.720,
    "args": {
      "External id": 950,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 330,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 330, "pid": 537, "tid": 537, "ts": 7429564826818.653,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826832.209, "dur": 3.916,
    "args": {
      "External id": 953,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 331,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 331, "pid": 537, "tid": 537, "ts": 7429564826832.209,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826864.874, "dur": 4.052,
    "args": {
      "External id": 955,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 332,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 332, "pid": 537, "tid": 537, "ts": 7429564826864.874,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826878.714, "dur": 5.069,
    "args": {
      "External id": 958,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 333,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 333, "pid": 537, "tid": 537, "ts": 7429564826878.714,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826893.313, "dur": 3.958,
    "args": {
      "External id": 959,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 334,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 334, "pid": 537, "tid": 537, "ts": 7429564826893.313,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826902.594, "dur": 4.879,
    "args": {
      "External id": 960,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 335,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 335, "pid": 537, "tid": 537, "ts": 7429564826902.594,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826913.601, "dur": 4.575,
    "args": {
      "External id": 961,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 336,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 336, "pid": 537, "tid": 537, "ts": 7429564826913.601,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826926.487, "dur": 3.407,
    "args": {
      "External id": 963,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 337,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 337, "pid": 537, "tid": 537, "ts": 7429564826926.487,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826970.344, "dur": 0.233,
    "args": {
      "External id": 971,      "cid": 370, "correlation": 338
    }
  },
  {
    "ph": "f", "id": 338, "pid": 537, "tid": 537, "ts": 7429564826970.344,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826974.122, "dur": 0.170,
    "args": {
      "External id": 971,      "cid": 370, "correlation": 339
    }
  },
  {
    "ph": "f", "id": 339, "pid": 537, "tid": 537, "ts": 7429564826974.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564826976.362, "dur": 0.149,
    "args": {
      "External id": 971,      "cid": 370, "correlation": 340
    }
  },
  {
    "ph": "f", "id": 340, "pid": 537, "tid": 537, "ts": 7429564826976.362,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564826978.464, "dur": 4.867,
    "args": {
      "External id": 971,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 341,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 341, "pid": 537, "tid": 537, "ts": 7429564826978.464,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827020.163, "dur": 0.182,
    "args": {
      "External id": 984,      "cid": 370, "correlation": 342
    }
  },
  {
    "ph": "f", "id": 342, "pid": 537, "tid": 537, "ts": 7429564827020.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827023.674, "dur": 0.153,
    "args": {
      "External id": 984,      "cid": 370, "correlation": 343
    }
  },
  {
    "ph": "f", "id": 343, "pid": 537, "tid": 537, "ts": 7429564827023.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827025.772, "dur": 0.149,
    "args": {
      "External id": 984,      "cid": 370, "correlation": 344
    }
  },
  {
    "ph": "f", "id": 344, "pid": 537, "tid": 537, "ts": 7429564827025.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827030.166, "dur": 4.374,
    "args": {
      "External id": 984,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 345,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 345, "pid": 537, "tid": 537, "ts": 7429564827030.166,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827035.064, "dur": 4.006,
    "args": {
      "External id": 984,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 346,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 346, "pid": 537, "tid": 537, "ts": 7429564827035.064,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827075.214, "dur": 0.185,
    "args": {
      "External id": 997,      "cid": 370, "correlation": 347
    }
  },
  {
    "ph": "f", "id": 347, "pid": 537, "tid": 537, "ts": 7429564827075.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827077.865, "dur": 0.140,
    "args": {
      "External id": 997,      "cid": 370, "correlation": 348
    }
  },
  {
    "ph": "f", "id": 348, "pid": 537, "tid": 537, "ts": 7429564827077.865,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827079.665, "dur": 0.152,
    "args": {
      "External id": 997,      "cid": 370, "correlation": 349
    }
  },
  {
    "ph": "f", "id": 349, "pid": 537, "tid": 537, "ts": 7429564827079.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827083.211, "dur": 4.532,
    "args": {
      "External id": 997,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 350,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 350, "pid": 537, "tid": 537, "ts": 7429564827083.211,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827088.037, "dur": 3.260,
    "args": {
      "External id": 997,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 351,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 351, "pid": 537, "tid": 537, "ts": 7429564827088.037,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827114.298, "dur": 4.827,
    "args": {
      "External id": 1007,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 352,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 352, "pid": 537, "tid": 537, "ts": 7429564827114.298,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827141.322, "dur": 4.585,
    "args": {
      "External id": 1012,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 353,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 353, "pid": 537, "tid": 537, "ts": 7429564827141.322,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827161.850, "dur": 6.983,
    "args": {
      "External id": 1013,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 354,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 354, "pid": 537, "tid": 537, "ts": 7429564827161.850,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827176.458, "dur": 3.801,
    "args": {
      "External id": 1014,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 355,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 355, "pid": 537, "tid": 537, "ts": 7429564827176.458,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827186.475, "dur": 5.293,
    "args": {
      "External id": 1015,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 356,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 356, "pid": 537, "tid": 537, "ts": 7429564827186.475,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827197.386, "dur": 3.289,
    "args": {
      "External id": 1016,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 357,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 357, "pid": 537, "tid": 537, "ts": 7429564827197.386,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827214.372, "dur": 3.712,
    "args": {
      "External id": 1021,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 358,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 358, "pid": 537, "tid": 537, "ts": 7429564827214.372,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827224.350, "dur": 4.865,
    "args": {
      "External id": 1022,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 359,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 359, "pid": 537, "tid": 537, "ts": 7429564827224.350,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827235.591, "dur": 3.906,
    "args": {
      "External id": 1023,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 360,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 360, "pid": 537, "tid": 537, "ts": 7429564827235.591,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827244.282, "dur": 3.482,
    "args": {
      "External id": 1024,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 361,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 361, "pid": 537, "tid": 537, "ts": 7429564827244.282,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827339.240, "dur": 4.512,
    "args": {
      "External id": 1041,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 362,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 362, "pid": 537, "tid": 537, "ts": 7429564827339.240,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827372.499, "dur": 3.736,
    "args": {
      "External id": 1059,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 363,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 363, "pid": 537, "tid": 537, "ts": 7429564827372.499,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827411.652, "dur": 0.368,
    "args": {
      "External id": 1072,      "cid": 370, "correlation": 364
    }
  },
  {
    "ph": "f", "id": 364, "pid": 537, "tid": 537, "ts": 7429564827411.652,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827416.458, "dur": 0.152,
    "args": {
      "External id": 1072,      "cid": 370, "correlation": 365
    }
  },
  {
    "ph": "f", "id": 365, "pid": 537, "tid": 537, "ts": 7429564827416.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827419.465, "dur": 0.152,
    "args": {
      "External id": 1072,      "cid": 370, "correlation": 366
    }
  },
  {
    "ph": "f", "id": 366, "pid": 537, "tid": 537, "ts": 7429564827419.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827422.587, "dur": 4.668,
    "args": {
      "External id": 1072,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 367,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 367, "pid": 537, "tid": 537, "ts": 7429564827422.587,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827439.863, "dur": 4.147,
    "args": {
      "External id": 1075,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 368,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 368, "pid": 537, "tid": 537, "ts": 7429564827439.863,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827457.958, "dur": 4.403,
    "args": {
      "External id": 1082,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 369,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 369, "pid": 537, "tid": 537, "ts": 7429564827457.958,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827473.822, "dur": 5.568,
    "args": {
      "External id": 1085,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 370,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 370, "pid": 537, "tid": 537, "ts": 7429564827473.822,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827522.513, "dur": 0.211,
    "args": {
      "External id": 1097,      "cid": 370, "correlation": 371
    }
  },
  {
    "ph": "f", "id": 371, "pid": 537, "tid": 537, "ts": 7429564827522.513,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827526.230, "dur": 0.146,
    "args": {
      "External id": 1097,      "cid": 370, "correlation": 372
    }
  },
  {
    "ph": "f", "id": 372, "pid": 537, "tid": 537, "ts": 7429564827526.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827529.187, "dur": 0.187,
    "args": {
      "External id": 1097,      "cid": 370, "correlation": 373
    }
  },
  {
    "ph": "f", "id": 373, "pid": 537, "tid": 537, "ts": 7429564827529.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827531.475, "dur": 5.398,
    "args": {
      "External id": 1097,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 374,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 374, "pid": 537, "tid": 537, "ts": 7429564827531.475,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827553.319, "dur": 3.692,
    "args": {
      "External id": 1106,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 375,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 375, "pid": 537, "tid": 537, "ts": 7429564827553.319,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827600.999, "dur": 0.202,
    "args": {
      "External id": 1116,      "cid": 370, "correlation": 376
    }
  },
  {
    "ph": "f", "id": 376, "pid": 537, "tid": 537, "ts": 7429564827600.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827604.004, "dur": 0.154,
    "args": {
      "External id": 1116,      "cid": 370, "correlation": 377
    }
  },
  {
    "ph": "f", "id": 377, "pid": 537, "tid": 537, "ts": 7429564827604.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827606.422, "dur": 0.194,
    "args": {
      "External id": 1116,      "cid": 370, "correlation": 378
    }
  },
  {
    "ph": "f", "id": 378, "pid": 537, "tid": 537, "ts": 7429564827606.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827608.311, "dur": 4.306,
    "args": {
      "External id": 1116,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 379,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 379, "pid": 537, "tid": 537, "ts": 7429564827608.311,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827625.042, "dur": 4.615,
    "args": {
      "External id": 1119,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 380,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 380, "pid": 537, "tid": 537, "ts": 7429564827625.042,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827647.558, "dur": 5.134,
    "args": {
      "External id": 1121,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 381,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 381, "pid": 537, "tid": 537, "ts": 7429564827647.558,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827663.163, "dur": 5.270,
    "args": {
      "External id": 1124,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 382,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 382, "pid": 537, "tid": 537, "ts": 7429564827663.163,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827677.483, "dur": 4.435,
    "args": {
      "External id": 1125,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 383,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 383, "pid": 537, "tid": 537, "ts": 7429564827677.483,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827687.822, "dur": 4.683,
    "args": {
      "External id": 1126,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 384,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 384, "pid": 537, "tid": 537, "ts": 7429564827687.822,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827699.149, "dur": 5.208,
    "args": {
      "External id": 1127,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 385,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 385, "pid": 537, "tid": 537, "ts": 7429564827699.149,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564827713.572, "dur": 0.307,
    "args": {
      "External id": 1129,      "cid": 290, "correlation": 386
    }
  },
  {
    "ph": "f", "id": 386, "pid": 537, "tid": 537, "ts": 7429564827713.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564827714.297, "dur": 61.042,
    "args": {
      "External id": 1129,            "bytes": 2097152, "cid": 108, "correlation": 387, "ptr": "0x100000002"
    }
  },
  {
    "ph": "s", "id": 387, "pid": 537, "tid": 537, "ts": 7429564827714.297,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827778.282, "dur": 3.971,
    "args": {
      "External id": 1129,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 388,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 388, "pid": 537, "tid": 537, "ts": 7429564827778.282,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827828.781, "dur": 0.311,
    "args": {
      "External id": 1137,      "cid": 370, "correlation": 389
    }
  },
  {
    "ph": "f", "id": 389, "pid": 537, "tid": 537, "ts": 7429564827828.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827833.337, "dur": 0.158,
    "args": {
      "External id": 1137,      "cid": 370, "correlation": 390
    }
  },
  {
    "ph": "f", "id": 390, "pid": 537, "tid": 537, "ts": 7429564827833.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827835.984, "dur": 0.150,
    "args": {
      "External id": 1137,      "cid": 370, "correlation": 391
    }
  },
  {
    "ph": "f", "id": 391, "pid": 537, "tid": 537, "ts": 7429564827835.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827844.849, "dur": 5.735,
    "args": {
      "External id": 1137,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 392,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 392, "pid": 537, "tid": 537, "ts": 7429564827844.849,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827850.978, "dur": 4.332,
    "args": {
      "External id": 1137,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 393,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 393, "pid": 537, "tid": 537, "ts": 7429564827850.978,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827873.951, "dur": 4.963,
    "args": {
      "External id": 1140,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 394,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 394, "pid": 537, "tid": 537, "ts": 7429564827873.951,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564827897.940, "dur": 0.213,
    "args": {
      "External id": 1148,      "cid": 290, "correlation": 395
    }
  },
  {
    "ph": "f", "id": 395, "pid": 537, "tid": 537, "ts": 7429564827897.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564827898.659, "dur": 32.473,
    "args": {
      "External id": 1148,            "bytes": 2097152, "cid": 108, "correlation": 396, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 396, "pid": 537, "tid": 537, "ts": 7429564827898.659,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827942.856, "dur": 0.206,
    "args": {
      "External id": 1148,      "cid": 370, "correlation": 397
    }
  },
  {
    "ph": "f", "id": 397, "pid": 537, "tid": 537, "ts": 7429564827942.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827946.949, "dur": 0.139,
    "args": {
      "External id": 1148,      "cid": 370, "correlation": 398
    }
  },
  {
    "ph": "f", "id": 398, "pid": 537, "tid": 537, "ts": 7429564827946.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564827949.711, "dur": 0.154,
    "args": {
      "External id": 1148,      "cid": 370, "correlation": 399
    }
  },
  {
    "ph": "f", "id": 399, "pid": 537, "tid": 537, "ts": 7429564827949.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827954.767, "dur": 5.082,
    "args": {
      "External id": 1148,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 400,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 400, "pid": 537, "tid": 537, "ts": 7429564827954.767,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827960.195, "dur": 3.920,
    "args": {
      "External id": 1148,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 401,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 401, "pid": 537, "tid": 537, "ts": 7429564827960.195,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564827974.787, "dur": 4.180,
    "args": {
      "External id": 1151,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 402,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 402, "pid": 537, "tid": 537, "ts": 7429564827974.787,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828004.535, "dur": 0.145,
    "args": {
      "External id": 1159,      "cid": 370, "correlation": 403
    }
  },
  {
    "ph": "f", "id": 403, "pid": 537, "tid": 537, "ts": 7429564828004.535,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828007.816, "dur": 0.167,
    "args": {
      "External id": 1159,      "cid": 370, "correlation": 404
    }
  },
  {
    "ph": "f", "id": 404, "pid": 537, "tid": 537, "ts": 7429564828007.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828009.975, "dur": 0.141,
    "args": {
      "External id": 1159,      "cid": 370, "correlation": 405
    }
  },
  {
    "ph": "f", "id": 405, "pid": 537, "tid": 537, "ts": 7429564828009.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828013.984, "dur": 4.537,
    "args": {
      "External id": 1159,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 406,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 406, "pid": 537, "tid": 537, "ts": 7429564828013.984,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828018.837, "dur": 3.160,
    "args": {
      "External id": 1159,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 407,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 407, "pid": 537, "tid": 537, "ts": 7429564828018.837,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828031.958, "dur": 4.590,
    "args": {
      "External id": 1162,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 408,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 408, "pid": 537, "tid": 537, "ts": 7429564828031.958,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828065.959, "dur": 4.856,
    "args": {
      "External id": 1164,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 409,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 409, "pid": 537, "tid": 537, "ts": 7429564828065.959,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828080.753, "dur": 6.038,
    "args": {
      "External id": 1167,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 410,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 410, "pid": 537, "tid": 537, "ts": 7429564828080.753,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828096.276, "dur": 3.884,
    "args": {
      "External id": 1168,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 411,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 411, "pid": 537, "tid": 537, "ts": 7429564828096.276,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828105.261, "dur": 3.812,
    "args": {
      "External id": 1169,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 412,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 412, "pid": 537, "tid": 537, "ts": 7429564828105.261,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828116.131, "dur": 4.697,
    "args": {
      "External id": 1170,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 413,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 413, "pid": 537, "tid": 537, "ts": 7429564828116.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828128.231, "dur": 5.165,
    "args": {
      "External id": 1172,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 414,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 414, "pid": 537, "tid": 537, "ts": 7429564828128.231,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564828164.216, "dur": 0.204,
    "args": {
      "External id": 1180,      "cid": 290, "correlation": 415
    }
  },
  {
    "ph": "f", "id": 415, "pid": 537, "tid": 537, "ts": 7429564828164.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564828164.712, "dur": 35.826,
    "args": {
      "External id": 1180,            "bytes": 2097152, "cid": 108, "correlation": 416, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 416, "pid": 537, "tid": 537, "ts": 7429564828164.712,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828213.717, "dur": 0.300,
    "args": {
      "External id": 1180,      "cid": 370, "correlation": 417
    }
  },
  {
    "ph": "f", "id": 417, "pid": 537, "tid": 537, "ts": 7429564828213.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828218.121, "dur": 0.154,
    "args": {
      "External id": 1180,      "cid": 370, "correlation": 418
    }
  },
  {
    "ph": "f", "id": 418, "pid": 537, "tid": 537, "ts": 7429564828218.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828220.687, "dur": 0.163,
    "args": {
      "External id": 1180,      "cid": 370, "correlation": 419
    }
  },
  {
    "ph": "f", "id": 419, "pid": 537, "tid": 537, "ts": 7429564828220.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828223.462, "dur": 5.609,
    "args": {
      "External id": 1180,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 420,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 420, "pid": 537, "tid": 537, "ts": 7429564828223.462,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828272.389, "dur": 0.166,
    "args": {
      "External id": 1193,      "cid": 370, "correlation": 421
    }
  },
  {
    "ph": "f", "id": 421, "pid": 537, "tid": 537, "ts": 7429564828272.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828275.777, "dur": 0.142,
    "args": {
      "External id": 1193,      "cid": 370, "correlation": 422
    }
  },
  {
    "ph": "f", "id": 422, "pid": 537, "tid": 537, "ts": 7429564828275.777,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828277.975, "dur": 0.159,
    "args": {
      "External id": 1193,      "cid": 370, "correlation": 423
    }
  },
  {
    "ph": "f", "id": 423, "pid": 537, "tid": 537, "ts": 7429564828277.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828283.136, "dur": 5.607,
    "args": {
      "External id": 1193,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 424,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 424, "pid": 537, "tid": 537, "ts": 7429564828283.136,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828289.173, "dur": 3.647,
    "args": {
      "External id": 1193,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 425,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 425, "pid": 537, "tid": 537, "ts": 7429564828289.173,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828331.862, "dur": 0.148,
    "args": {
      "External id": 1206,      "cid": 370, "correlation": 426
    }
  },
  {
    "ph": "f", "id": 426, "pid": 537, "tid": 537, "ts": 7429564828331.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828334.495, "dur": 0.145,
    "args": {
      "External id": 1206,      "cid": 370, "correlation": 427
    }
  },
  {
    "ph": "f", "id": 427, "pid": 537, "tid": 537, "ts": 7429564828334.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828336.502, "dur": 0.154,
    "args": {
      "External id": 1206,      "cid": 370, "correlation": 428
    }
  },
  {
    "ph": "f", "id": 428, "pid": 537, "tid": 537, "ts": 7429564828336.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828340.074, "dur": 4.767,
    "args": {
      "External id": 1206,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 429,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 429, "pid": 537, "tid": 537, "ts": 7429564828340.074,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828345.359, "dur": 3.430,
    "args": {
      "External id": 1206,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 430,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 430, "pid": 537, "tid": 537, "ts": 7429564828345.359,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828372.932, "dur": 5.362,
    "args": {
      "External id": 1216,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 431,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 431, "pid": 537, "tid": 537, "ts": 7429564828372.932,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828400.146, "dur": 5.240,
    "args": {
      "External id": 1221,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 432,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 432, "pid": 537, "tid": 537, "ts": 7429564828400.146,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828419.272, "dur": 7.279,
    "args": {
      "External id": 1222,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 433,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 433, "pid": 537, "tid": 537, "ts": 7429564828419.272,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828433.953, "dur": 3.600,
    "args": {
      "External id": 1223,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 434,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 434, "pid": 537, "tid": 537, "ts": 7429564828433.953,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828444.169, "dur": 4.514,
    "args": {
      "External id": 1224,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 435,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 435, "pid": 537, "tid": 537, "ts": 7429564828444.169,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828454.372, "dur": 3.475,
    "args": {
      "External id": 1225,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 436,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 436, "pid": 537, "tid": 537, "ts": 7429564828454.372,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828471.883, "dur": 4.305,
    "args": {
      "External id": 1230,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 437,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 437, "pid": 537, "tid": 537, "ts": 7429564828471.883,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828482.481, "dur": 5.049,
    "args": {
      "External id": 1231,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 438,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 438, "pid": 537, "tid": 537, "ts": 7429564828482.481,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828493.480, "dur": 4.084,
    "args": {
      "External id": 1232,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 439,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 439, "pid": 537, "tid": 537, "ts": 7429564828493.480,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828502.409, "dur": 3.269,
    "args": {
      "External id": 1233,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 440,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 440, "pid": 537, "tid": 537, "ts": 7429564828502.409,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828563.682, "dur": 4.905,
    "args": {
      "External id": 1250,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 441,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 441, "pid": 537, "tid": 537, "ts": 7429564828563.682,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828597.634, "dur": 3.465,
    "args": {
      "External id": 1268,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 442,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 442, "pid": 537, "tid": 537, "ts": 7429564828597.634,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828634.986, "dur": 0.306,
    "args": {
      "External id": 1281,      "cid": 370, "correlation": 443
    }
  },
  {
    "ph": "f", "id": 443, "pid": 537, "tid": 537, "ts": 7429564828634.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828639.457, "dur": 0.159,
    "args": {
      "External id": 1281,      "cid": 370, "correlation": 444
    }
  },
  {
    "ph": "f", "id": 444, "pid": 537, "tid": 537, "ts": 7429564828639.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828641.775, "dur": 0.165,
    "args": {
      "External id": 1281,      "cid": 370, "correlation": 445
    }
  },
  {
    "ph": "f", "id": 445, "pid": 537, "tid": 537, "ts": 7429564828641.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828644.325, "dur": 5.427,
    "args": {
      "External id": 1281,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 446,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 446, "pid": 537, "tid": 537, "ts": 7429564828644.325,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828661.261, "dur": 3.915,
    "args": {
      "External id": 1284,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 447,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 447, "pid": 537, "tid": 537, "ts": 7429564828661.261,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828679.161, "dur": 3.570,
    "args": {
      "External id": 1291,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 448,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 448, "pid": 537, "tid": 537, "ts": 7429564828679.161,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828693.922, "dur": 5.383,
    "args": {
      "External id": 1294,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 449,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 449, "pid": 537, "tid": 537, "ts": 7429564828693.922,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828741.556, "dur": 0.204,
    "args": {
      "External id": 1306,      "cid": 370, "correlation": 450
    }
  },
  {
    "ph": "f", "id": 450, "pid": 537, "tid": 537, "ts": 7429564828741.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828745.019, "dur": 0.148,
    "args": {
      "External id": 1306,      "cid": 370, "correlation": 451
    }
  },
  {
    "ph": "f", "id": 451, "pid": 537, "tid": 537, "ts": 7429564828745.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828747.933, "dur": 0.142,
    "args": {
      "External id": 1306,      "cid": 370, "correlation": 452
    }
  },
  {
    "ph": "f", "id": 452, "pid": 537, "tid": 537, "ts": 7429564828747.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828749.924, "dur": 4.528,
    "args": {
      "External id": 1306,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 453,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 453, "pid": 537, "tid": 537, "ts": 7429564828749.924,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828770.662, "dur": 4.385,
    "args": {
      "External id": 1315,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 454,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 454, "pid": 537, "tid": 537, "ts": 7429564828770.662,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828815.396, "dur": 0.203,
    "args": {
      "External id": 1325,      "cid": 370, "correlation": 455
    }
  },
  {
    "ph": "f", "id": 455, "pid": 537, "tid": 537, "ts": 7429564828815.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828818.416, "dur": 0.154,
    "args": {
      "External id": 1325,      "cid": 370, "correlation": 456
    }
  },
  {
    "ph": "f", "id": 456, "pid": 537, "tid": 537, "ts": 7429564828818.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564828820.697, "dur": 0.151,
    "args": {
      "External id": 1325,      "cid": 370, "correlation": 457
    }
  },
  {
    "ph": "f", "id": 457, "pid": 537, "tid": 537, "ts": 7429564828820.697,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828822.654, "dur": 5.285,
    "args": {
      "External id": 1325,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 458,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 458, "pid": 537, "tid": 537, "ts": 7429564828822.654,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828839.605, "dur": 4.050,
    "args": {
      "External id": 1328,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 459,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 459, "pid": 537, "tid": 537, "ts": 7429564828839.605,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828861.586, "dur": 4.147,
    "args": {
      "External id": 1330,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 460,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 460, "pid": 537, "tid": 537, "ts": 7429564828861.586,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828876.154, "dur": 6.045,
    "args": {
      "External id": 1333,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 461,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 461, "pid": 537, "tid": 537, "ts": 7429564828876.154,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828890.453, "dur": 4.377,
    "args": {
      "External id": 1334,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 462,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 462, "pid": 537, "tid": 537, "ts": 7429564828890.453,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828900.092, "dur": 4.363,
    "args": {
      "External id": 1335,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 463,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 463, "pid": 537, "tid": 537, "ts": 7429564828900.092,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828913.074, "dur": 5.142,
    "args": {
      "External id": 1336,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 464,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 464, "pid": 537, "tid": 537, "ts": 7429564828913.074,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564828926.924, "dur": 3.461,
    "args": {
      "External id": 1338,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 465,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 465, "pid": 537, "tid": 537, "ts": 7429564828926.924,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564828957.038, "dur": 0.380,
    "args": {
      "External id": 1346,      "cid": 290, "correlation": 466
    }
  },
  {
    "ph": "f", "id": 466, "pid": 537, "tid": 537, "ts": 7429564828957.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564828957.758, "dur": 52.700,
    "args": {
      "External id": 1346,            "bytes": 2097152, "cid": 108, "correlation": 467, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 467, "pid": 537, "tid": 537, "ts": 7429564828957.758,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829024.179, "dur": 0.301,
    "args": {
      "External id": 1346,      "cid": 370, "correlation": 468
    }
  },
  {
    "ph": "f", "id": 468, "pid": 537, "tid": 537, "ts": 7429564829024.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829028.568, "dur": 0.172,
    "args": {
      "External id": 1346,      "cid": 370, "correlation": 469
    }
  },
  {
    "ph": "f", "id": 469, "pid": 537, "tid": 537, "ts": 7429564829028.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829031.140, "dur": 0.160,
    "args": {
      "External id": 1346,      "cid": 370, "correlation": 470
    }
  },
  {
    "ph": "f", "id": 470, "pid": 537, "tid": 537, "ts": 7429564829031.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829039.181, "dur": 5.872,
    "args": {
      "External id": 1346,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 471,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 471, "pid": 537, "tid": 537, "ts": 7429564829039.181,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829045.489, "dur": 4.173,
    "args": {
      "External id": 1346,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 472,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 472, "pid": 537, "tid": 537, "ts": 7429564829045.489,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829071.261, "dur": 4.722,
    "args": {
      "External id": 1349,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 473,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 473, "pid": 537, "tid": 537, "ts": 7429564829071.261,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564829097.601, "dur": 0.284,
    "args": {
      "External id": 1357,      "cid": 290, "correlation": 474
    }
  },
  {
    "ph": "f", "id": 474, "pid": 537, "tid": 537, "ts": 7429564829097.601,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564829098.184, "dur": 30.619,
    "args": {
      "External id": 1357,            "bytes": 2097152, "cid": 108, "correlation": 475, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 475, "pid": 537, "tid": 537, "ts": 7429564829098.184,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829140.398, "dur": 0.210,
    "args": {
      "External id": 1357,      "cid": 370, "correlation": 476
    }
  },
  {
    "ph": "f", "id": 476, "pid": 537, "tid": 537, "ts": 7429564829140.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829143.943, "dur": 0.164,
    "args": {
      "External id": 1357,      "cid": 370, "correlation": 477
    }
  },
  {
    "ph": "f", "id": 477, "pid": 537, "tid": 537, "ts": 7429564829143.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829146.481, "dur": 0.164,
    "args": {
      "External id": 1357,      "cid": 370, "correlation": 478
    }
  },
  {
    "ph": "f", "id": 478, "pid": 537, "tid": 537, "ts": 7429564829146.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829151.001, "dur": 5.864,
    "args": {
      "External id": 1357,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 479,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 479, "pid": 537, "tid": 537, "ts": 7429564829151.001,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829157.178, "dur": 4.695,
    "args": {
      "External id": 1357,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 480,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 480, "pid": 537, "tid": 537, "ts": 7429564829157.178,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829172.638, "dur": 4.383,
    "args": {
      "External id": 1360,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 481,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 481, "pid": 537, "tid": 537, "ts": 7429564829172.638,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829207.665, "dur": 0.173,
    "args": {
      "External id": 1368,      "cid": 370, "correlation": 482
    }
  },
  {
    "ph": "f", "id": 482, "pid": 537, "tid": 537, "ts": 7429564829207.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829210.995, "dur": 0.140,
    "args": {
      "External id": 1368,      "cid": 370, "correlation": 483
    }
  },
  {
    "ph": "f", "id": 483, "pid": 537, "tid": 537, "ts": 7429564829210.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829213.286, "dur": 0.154,
    "args": {
      "External id": 1368,      "cid": 370, "correlation": 484
    }
  },
  {
    "ph": "f", "id": 484, "pid": 537, "tid": 537, "ts": 7429564829213.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829217.080, "dur": 4.841,
    "args": {
      "External id": 1368,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 485,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 485, "pid": 537, "tid": 537, "ts": 7429564829217.080,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829222.411, "dur": 3.925,
    "args": {
      "External id": 1368,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 486,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 486, "pid": 537, "tid": 537, "ts": 7429564829222.411,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829236.527, "dur": 4.326,
    "args": {
      "External id": 1371,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 487,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 487, "pid": 537, "tid": 537, "ts": 7429564829236.527,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829269.282, "dur": 4.430,
    "args": {
      "External id": 1373,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 488,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 488, "pid": 537, "tid": 537, "ts": 7429564829269.282,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829283.782, "dur": 6.338,
    "args": {
      "External id": 1376,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 489,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 489, "pid": 537, "tid": 537, "ts": 7429564829283.782,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829299.240, "dur": 4.351,
    "args": {
      "External id": 1377,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 490,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 490, "pid": 537, "tid": 537, "ts": 7429564829299.240,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829309.101, "dur": 4.048,
    "args": {
      "External id": 1378,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 491,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 491, "pid": 537, "tid": 537, "ts": 7429564829309.101,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829320.358, "dur": 4.293,
    "args": {
      "External id": 1379,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 492,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 492, "pid": 537, "tid": 537, "ts": 7429564829320.358,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829331.906, "dur": 3.829,
    "args": {
      "External id": 1381,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 493,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 493, "pid": 537, "tid": 537, "ts": 7429564829331.906,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829376.979, "dur": 0.299,
    "args": {
      "External id": 1389,      "cid": 370, "correlation": 494
    }
  },
  {
    "ph": "f", "id": 494, "pid": 537, "tid": 537, "ts": 7429564829376.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829380.525, "dur": 0.162,
    "args": {
      "External id": 1389,      "cid": 370, "correlation": 495
    }
  },
  {
    "ph": "f", "id": 495, "pid": 537, "tid": 537, "ts": 7429564829380.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829382.957, "dur": 0.269,
    "args": {
      "External id": 1389,      "cid": 370, "correlation": 496
    }
  },
  {
    "ph": "f", "id": 496, "pid": 537, "tid": 537, "ts": 7429564829382.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829385.185, "dur": 5.683,
    "args": {
      "External id": 1389,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 497,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 497, "pid": 537, "tid": 537, "ts": 7429564829385.185,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829428.969, "dur": 0.193,
    "args": {
      "External id": 1402,      "cid": 370, "correlation": 498
    }
  },
  {
    "ph": "f", "id": 498, "pid": 537, "tid": 537, "ts": 7429564829428.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829431.870, "dur": 0.146,
    "args": {
      "External id": 1402,      "cid": 370, "correlation": 499
    }
  },
  {
    "ph": "f", "id": 499, "pid": 537, "tid": 537, "ts": 7429564829431.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829434.135, "dur": 0.148,
    "args": {
      "External id": 1402,      "cid": 370, "correlation": 500
    }
  },
  {
    "ph": "f", "id": 500, "pid": 537, "tid": 537, "ts": 7429564829434.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829438.340, "dur": 5.589,
    "args": {
      "External id": 1402,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 501,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 501, "pid": 537, "tid": 537, "ts": 7429564829438.340,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829444.494, "dur": 4.228,
    "args": {
      "External id": 1402,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 502,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 502, "pid": 537, "tid": 537, "ts": 7429564829444.494,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829484.514, "dur": 0.168,
    "args": {
      "External id": 1415,      "cid": 370, "correlation": 503
    }
  },
  {
    "ph": "f", "id": 503, "pid": 537, "tid": 537, "ts": 7429564829484.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829487.176, "dur": 0.141,
    "args": {
      "External id": 1415,      "cid": 370, "correlation": 504
    }
  },
  {
    "ph": "f", "id": 504, "pid": 537, "tid": 537, "ts": 7429564829487.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829488.960, "dur": 0.139,
    "args": {
      "External id": 1415,      "cid": 370, "correlation": 505
    }
  },
  {
    "ph": "f", "id": 505, "pid": 537, "tid": 537, "ts": 7429564829488.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829492.564, "dur": 4.179,
    "args": {
      "External id": 1415,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 506,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 506, "pid": 537, "tid": 537, "ts": 7429564829492.564,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829497.037, "dur": 4.079,
    "args": {
      "External id": 1415,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 507,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 507, "pid": 537, "tid": 537, "ts": 7429564829497.037,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829525.487, "dur": 4.353,
    "args": {
      "External id": 1425,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 508,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 508, "pid": 537, "tid": 537, "ts": 7429564829525.487,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829552.272, "dur": 4.659,
    "args": {
      "External id": 1430,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 509,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 509, "pid": 537, "tid": 537, "ts": 7429564829552.272,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829571.352, "dur": 7.692,
    "args": {
      "External id": 1431,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 510,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 510, "pid": 537, "tid": 537, "ts": 7429564829571.352,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564829587.133, "dur": 0.270,
    "args": {
      "External id": 1432,      "cid": 290, "correlation": 511
    }
  },
  {
    "ph": "f", "id": 511, "pid": 537, "tid": 537, "ts": 7429564829587.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564829587.685, "dur": 36.012,
    "args": {
      "External id": 1432,            "bytes": 2097152, "cid": 108, "correlation": 512, "ptr": "0x2"
    }
  },
  {
    "ph": "s", "id": 512, "pid": 537, "tid": 537, "ts": 7429564829587.685,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829625.821, "dur": 4.350,
    "args": {
      "External id": 1432,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 513,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 513, "pid": 537, "tid": 537, "ts": 7429564829625.821,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829640.479, "dur": 4.405,
    "args": {
      "External id": 1433,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 514,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 514, "pid": 537, "tid": 537, "ts": 7429564829640.479,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829650.462, "dur": 3.472,
    "args": {
      "External id": 1434,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 515,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 515, "pid": 537, "tid": 537, "ts": 7429564829650.462,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829669.007, "dur": 3.629,
    "args": {
      "External id": 1439,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 516,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 516, "pid": 537, "tid": 537, "ts": 7429564829669.007,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829680.396, "dur": 4.790,
    "args": {
      "External id": 1440,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 517,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 517, "pid": 537, "tid": 537, "ts": 7429564829680.396,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829691.836, "dur": 3.798,
    "args": {
      "External id": 1441,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 518,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 518, "pid": 537, "tid": 537, "ts": 7429564829691.836,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829699.988, "dur": 3.653,
    "args": {
      "External id": 1442,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 519,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 519, "pid": 537, "tid": 537, "ts": 7429564829699.988,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829764.895, "dur": 4.940,
    "args": {
      "External id": 1459,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 520,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 520, "pid": 537, "tid": 537, "ts": 7429564829764.895,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829800.074, "dur": 3.609,
    "args": {
      "External id": 1477,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 521,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 521, "pid": 537, "tid": 537, "ts": 7429564829800.074,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829840.361, "dur": 0.316,
    "args": {
      "External id": 1490,      "cid": 370, "correlation": 522
    }
  },
  {
    "ph": "f", "id": 522, "pid": 537, "tid": 537, "ts": 7429564829840.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829845.481, "dur": 0.186,
    "args": {
      "External id": 1490,      "cid": 370, "correlation": 523
    }
  },
  {
    "ph": "f", "id": 523, "pid": 537, "tid": 537, "ts": 7429564829845.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829848.810, "dur": 0.160,
    "args": {
      "External id": 1490,      "cid": 370, "correlation": 524
    }
  },
  {
    "ph": "f", "id": 524, "pid": 537, "tid": 537, "ts": 7429564829848.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829851.931, "dur": 6.314,
    "args": {
      "External id": 1490,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 525,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 525, "pid": 537, "tid": 537, "ts": 7429564829851.931,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829870.765, "dur": 4.216,
    "args": {
      "External id": 1493,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 526,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 526, "pid": 537, "tid": 537, "ts": 7429564829870.765,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829890.423, "dur": 4.827,
    "args": {
      "External id": 1500,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 527,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 527, "pid": 537, "tid": 537, "ts": 7429564829890.423,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829907.497, "dur": 6.984,
    "args": {
      "External id": 1503,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 528,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 528, "pid": 537, "tid": 537, "ts": 7429564829907.497,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829958.553, "dur": 0.220,
    "args": {
      "External id": 1515,      "cid": 370, "correlation": 529
    }
  },
  {
    "ph": "f", "id": 529, "pid": 537, "tid": 537, "ts": 7429564829958.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829961.823, "dur": 0.161,
    "args": {
      "External id": 1515,      "cid": 370, "correlation": 530
    }
  },
  {
    "ph": "f", "id": 530, "pid": 537, "tid": 537, "ts": 7429564829961.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564829964.679, "dur": 0.152,
    "args": {
      "External id": 1515,      "cid": 370, "correlation": 531
    }
  },
  {
    "ph": "f", "id": 531, "pid": 537, "tid": 537, "ts": 7429564829964.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829966.873, "dur": 5.285,
    "args": {
      "External id": 1515,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 532,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 532, "pid": 537, "tid": 537, "ts": 7429564829966.873,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564829989.045, "dur": 4.056,
    "args": {
      "External id": 1524,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 533,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 533, "pid": 537, "tid": 537, "ts": 7429564829989.045,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830034.524, "dur": 0.278,
    "args": {
      "External id": 1534,      "cid": 370, "correlation": 534
    }
  },
  {
    "ph": "f", "id": 534, "pid": 537, "tid": 537, "ts": 7429564830034.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830037.545, "dur": 0.145,
    "args": {
      "External id": 1534,      "cid": 370, "correlation": 535
    }
  },
  {
    "ph": "f", "id": 535, "pid": 537, "tid": 537, "ts": 7429564830037.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830039.839, "dur": 0.219,
    "args": {
      "External id": 1534,      "cid": 370, "correlation": 536
    }
  },
  {
    "ph": "f", "id": 536, "pid": 537, "tid": 537, "ts": 7429564830039.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830041.643, "dur": 4.683,
    "args": {
      "External id": 1534,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 537,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 537, "pid": 537, "tid": 537, "ts": 7429564830041.643,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830113.546, "dur": 5.477,
    "args": {
      "External id": 1537,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 538,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 538, "pid": 537, "tid": 537, "ts": 7429564830113.546,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830141.897, "dur": 5.309,
    "args": {
      "External id": 1539,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 539,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 539, "pid": 537, "tid": 537, "ts": 7429564830141.897,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830158.756, "dur": 6.767,
    "args": {
      "External id": 1542,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 540,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 540, "pid": 537, "tid": 537, "ts": 7429564830158.756,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830174.449, "dur": 4.052,
    "args": {
      "External id": 1543,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 541,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 541, "pid": 537, "tid": 537, "ts": 7429564830174.449,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830185.519, "dur": 4.516,
    "args": {
      "External id": 1544,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 542,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 542, "pid": 537, "tid": 537, "ts": 7429564830185.519,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830197.289, "dur": 4.002,
    "args": {
      "External id": 1545,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 543,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 543, "pid": 537, "tid": 537, "ts": 7429564830197.289,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830209.506, "dur": 5.246,
    "args": {
      "External id": 1547,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 544,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 544, "pid": 537, "tid": 537, "ts": 7429564830209.506,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830253.295, "dur": 0.223,
    "args": {
      "External id": 1555,      "cid": 370, "correlation": 545
    }
  },
  {
    "ph": "f", "id": 545, "pid": 537, "tid": 537, "ts": 7429564830253.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830257.443, "dur": 0.159,
    "args": {
      "External id": 1555,      "cid": 370, "correlation": 546
    }
  },
  {
    "ph": "f", "id": 546, "pid": 537, "tid": 537, "ts": 7429564830257.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830259.932, "dur": 0.166,
    "args": {
      "External id": 1555,      "cid": 370, "correlation": 547
    }
  },
  {
    "ph": "f", "id": 547, "pid": 537, "tid": 537, "ts": 7429564830259.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830268.995, "dur": 6.167,
    "args": {
      "External id": 1555,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 548,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 548, "pid": 537, "tid": 537, "ts": 7429564830268.995,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830275.603, "dur": 3.961,
    "args": {
      "External id": 1555,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 549,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 549, "pid": 537, "tid": 537, "ts": 7429564830275.603,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564830297.794, "dur": 0.312,
    "args": {
      "External id": 1558,      "cid": 290, "correlation": 550
    }
  },
  {
    "ph": "f", "id": 550, "pid": 537, "tid": 537, "ts": 7429564830297.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564830298.611, "dur": 56.557,
    "args": {
      "External id": 1558,            "bytes": 2097152, "cid": 108, "correlation": 551, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 551, "pid": 537, "tid": 537, "ts": 7429564830298.611,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830358.292, "dur": 4.439,
    "args": {
      "External id": 1558,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 552,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 552, "pid": 537, "tid": 537, "ts": 7429564830358.292,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830396.599, "dur": 0.204,
    "args": {
      "External id": 1566,      "cid": 370, "correlation": 553
    }
  },
  {
    "ph": "f", "id": 553, "pid": 537, "tid": 537, "ts": 7429564830396.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830399.672, "dur": 0.150,
    "args": {
      "External id": 1566,      "cid": 370, "correlation": 554
    }
  },
  {
    "ph": "f", "id": 554, "pid": 537, "tid": 537, "ts": 7429564830399.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830401.987, "dur": 0.143,
    "args": {
      "External id": 1566,      "cid": 370, "correlation": 555
    }
  },
  {
    "ph": "f", "id": 555, "pid": 537, "tid": 537, "ts": 7429564830401.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830407.082, "dur": 5.784,
    "args": {
      "External id": 1566,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 556,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 556, "pid": 537, "tid": 537, "ts": 7429564830407.082,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830413.164, "dur": 3.707,
    "args": {
      "External id": 1566,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 557,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 557, "pid": 537, "tid": 537, "ts": 7429564830413.164,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564830426.227, "dur": 0.180,
    "args": {
      "External id": 1569,      "cid": 290, "correlation": 558
    }
  },
  {
    "ph": "f", "id": 558, "pid": 537, "tid": 537, "ts": 7429564830426.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564830426.681, "dur": 29.354,
    "args": {
      "External id": 1569,            "bytes": 2097152, "cid": 108, "correlation": 559, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 559, "pid": 537, "tid": 537, "ts": 7429564830426.681,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830458.230, "dur": 4.097,
    "args": {
      "External id": 1569,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 560,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 560, "pid": 537, "tid": 537, "ts": 7429564830458.230,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830491.837, "dur": 0.219,
    "args": {
      "External id": 1577,      "cid": 370, "correlation": 561
    }
  },
  {
    "ph": "f", "id": 561, "pid": 537, "tid": 537, "ts": 7429564830491.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830495.255, "dur": 0.149,
    "args": {
      "External id": 1577,      "cid": 370, "correlation": 562
    }
  },
  {
    "ph": "f", "id": 562, "pid": 537, "tid": 537, "ts": 7429564830495.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830497.505, "dur": 0.152,
    "args": {
      "External id": 1577,      "cid": 370, "correlation": 563
    }
  },
  {
    "ph": "f", "id": 563, "pid": 537, "tid": 537, "ts": 7429564830497.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830502.004, "dur": 5.568,
    "args": {
      "External id": 1577,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 564,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 564, "pid": 537, "tid": 537, "ts": 7429564830502.004,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830507.869, "dur": 4.026,
    "args": {
      "External id": 1577,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 565,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 565, "pid": 537, "tid": 537, "ts": 7429564830507.869,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830522.310, "dur": 3.987,
    "args": {
      "External id": 1580,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 566,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 566, "pid": 537, "tid": 537, "ts": 7429564830522.310,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830555.074, "dur": 3.689,
    "args": {
      "External id": 1582,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 567,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 567, "pid": 537, "tid": 537, "ts": 7429564830555.074,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830568.423, "dur": 4.409,
    "args": {
      "External id": 1585,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 568,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 568, "pid": 537, "tid": 537, "ts": 7429564830568.423,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830582.291, "dur": 4.642,
    "args": {
      "External id": 1586,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 569,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 569, "pid": 537, "tid": 537, "ts": 7429564830582.291,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830592.482, "dur": 4.860,
    "args": {
      "External id": 1587,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 570,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 570, "pid": 537, "tid": 537, "ts": 7429564830592.482,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830604.499, "dur": 4.307,
    "args": {
      "External id": 1588,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 571,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 571, "pid": 537, "tid": 537, "ts": 7429564830604.499,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830616.303, "dur": 4.970,
    "args": {
      "External id": 1590,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 572,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 572, "pid": 537, "tid": 537, "ts": 7429564830616.303,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830664.009, "dur": 0.196,
    "args": {
      "External id": 1598,      "cid": 370, "correlation": 573
    }
  },
  {
    "ph": "f", "id": 573, "pid": 537, "tid": 537, "ts": 7429564830664.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830667.175, "dur": 0.183,
    "args": {
      "External id": 1598,      "cid": 370, "correlation": 574
    }
  },
  {
    "ph": "f", "id": 574, "pid": 537, "tid": 537, "ts": 7429564830667.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830669.342, "dur": 0.154,
    "args": {
      "External id": 1598,      "cid": 370, "correlation": 575
    }
  },
  {
    "ph": "f", "id": 575, "pid": 537, "tid": 537, "ts": 7429564830669.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830671.471, "dur": 4.786,
    "args": {
      "External id": 1598,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 576,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 576, "pid": 537, "tid": 537, "ts": 7429564830671.471,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830715.334, "dur": 0.184,
    "args": {
      "External id": 1611,      "cid": 370, "correlation": 577
    }
  },
  {
    "ph": "f", "id": 577, "pid": 537, "tid": 537, "ts": 7429564830715.334,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830718.801, "dur": 0.163,
    "args": {
      "External id": 1611,      "cid": 370, "correlation": 578
    }
  },
  {
    "ph": "f", "id": 578, "pid": 537, "tid": 537, "ts": 7429564830718.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830721.319, "dur": 0.149,
    "args": {
      "External id": 1611,      "cid": 370, "correlation": 579
    }
  },
  {
    "ph": "f", "id": 579, "pid": 537, "tid": 537, "ts": 7429564830721.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830725.575, "dur": 4.727,
    "args": {
      "External id": 1611,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 580,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 580, "pid": 537, "tid": 537, "ts": 7429564830725.575,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830730.758, "dur": 5.032,
    "args": {
      "External id": 1611,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 581,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 581, "pid": 537, "tid": 537, "ts": 7429564830730.758,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830772.760, "dur": 0.186,
    "args": {
      "External id": 1624,      "cid": 370, "correlation": 582
    }
  },
  {
    "ph": "f", "id": 582, "pid": 537, "tid": 537, "ts": 7429564830772.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830775.488, "dur": 0.149,
    "args": {
      "External id": 1624,      "cid": 370, "correlation": 583
    }
  },
  {
    "ph": "f", "id": 583, "pid": 537, "tid": 537, "ts": 7429564830775.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564830777.600, "dur": 0.154,
    "args": {
      "External id": 1624,      "cid": 370, "correlation": 584
    }
  },
  {
    "ph": "f", "id": 584, "pid": 537, "tid": 537, "ts": 7429564830777.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830781.103, "dur": 5.233,
    "args": {
      "External id": 1624,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 585,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 585, "pid": 537, "tid": 537, "ts": 7429564830781.103,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830786.864, "dur": 3.843,
    "args": {
      "External id": 1624,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 586,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 586, "pid": 537, "tid": 537, "ts": 7429564830786.864,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830814.007, "dur": 3.670,
    "args": {
      "External id": 1634,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 587,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 587, "pid": 537, "tid": 537, "ts": 7429564830814.007,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830841.338, "dur": 4.241,
    "args": {
      "External id": 1639,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 588,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 588, "pid": 537, "tid": 537, "ts": 7429564830841.338,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830860.855, "dur": 7.014,
    "args": {
      "External id": 1640,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 589,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 589, "pid": 537, "tid": 537, "ts": 7429564830860.855,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830875.600, "dur": 4.009,
    "args": {
      "External id": 1641,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 590,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 590, "pid": 537, "tid": 537, "ts": 7429564830875.600,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830884.685, "dur": 4.920,
    "args": {
      "External id": 1642,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 591,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 591, "pid": 537, "tid": 537, "ts": 7429564830884.685,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830895.926, "dur": 4.680,
    "args": {
      "External id": 1643,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 592,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 592, "pid": 537, "tid": 537, "ts": 7429564830895.926,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830912.314, "dur": 3.805,
    "args": {
      "External id": 1648,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 593,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 593, "pid": 537, "tid": 537, "ts": 7429564830912.314,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830923.116, "dur": 5.222,
    "args": {
      "External id": 1649,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 594,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 594, "pid": 537, "tid": 537, "ts": 7429564830923.116,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830934.088, "dur": 3.208,
    "args": {
      "External id": 1650,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 595,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 595, "pid": 537, "tid": 537, "ts": 7429564830934.088,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564830941.876, "dur": 6.618,
    "args": {
      "External id": 1651,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 596,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 596, "pid": 537, "tid": 537, "ts": 7429564830941.876,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831008.862, "dur": 4.875,
    "args": {
      "External id": 1668,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 597,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 597, "pid": 537, "tid": 537, "ts": 7429564831008.862,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831043.922, "dur": 4.002,
    "args": {
      "External id": 1686,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 598,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 598, "pid": 537, "tid": 537, "ts": 7429564831043.922,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831085.468, "dur": 0.229,
    "args": {
      "External id": 1699,      "cid": 370, "correlation": 599
    }
  },
  {
    "ph": "f", "id": 599, "pid": 537, "tid": 537, "ts": 7429564831085.468,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831089.377, "dur": 0.167,
    "args": {
      "External id": 1699,      "cid": 370, "correlation": 600
    }
  },
  {
    "ph": "f", "id": 600, "pid": 537, "tid": 537, "ts": 7429564831089.377,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831091.632, "dur": 0.153,
    "args": {
      "External id": 1699,      "cid": 370, "correlation": 601
    }
  },
  {
    "ph": "f", "id": 601, "pid": 537, "tid": 537, "ts": 7429564831091.632,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831093.933, "dur": 4.760,
    "args": {
      "External id": 1699,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 602,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 602, "pid": 537, "tid": 537, "ts": 7429564831093.933,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831110.172, "dur": 5.205,
    "args": {
      "External id": 1702,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 603,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 603, "pid": 537, "tid": 537, "ts": 7429564831110.172,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831130.851, "dur": 3.530,
    "args": {
      "External id": 1709,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 604,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 604, "pid": 537, "tid": 537, "ts": 7429564831130.851,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831146.380, "dur": 5.221,
    "args": {
      "External id": 1712,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 605,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 605, "pid": 537, "tid": 537, "ts": 7429564831146.380,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831192.522, "dur": 0.290,
    "args": {
      "External id": 1724,      "cid": 370, "correlation": 606
    }
  },
  {
    "ph": "f", "id": 606, "pid": 537, "tid": 537, "ts": 7429564831192.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831196.881, "dur": 0.145,
    "args": {
      "External id": 1724,      "cid": 370, "correlation": 607
    }
  },
  {
    "ph": "f", "id": 607, "pid": 537, "tid": 537, "ts": 7429564831196.881,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831199.814, "dur": 0.159,
    "args": {
      "External id": 1724,      "cid": 370, "correlation": 608
    }
  },
  {
    "ph": "f", "id": 608, "pid": 537, "tid": 537, "ts": 7429564831199.814,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831202.212, "dur": 5.163,
    "args": {
      "External id": 1724,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 609,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 609, "pid": 537, "tid": 537, "ts": 7429564831202.212,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831223.996, "dur": 4.647,
    "args": {
      "External id": 1733,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 610,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 610, "pid": 537, "tid": 537, "ts": 7429564831223.996,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831270.664, "dur": 0.236,
    "args": {
      "External id": 1743,      "cid": 370, "correlation": 611
    }
  },
  {
    "ph": "f", "id": 611, "pid": 537, "tid": 537, "ts": 7429564831270.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831273.882, "dur": 0.169,
    "args": {
      "External id": 1743,      "cid": 370, "correlation": 612
    }
  },
  {
    "ph": "f", "id": 612, "pid": 537, "tid": 537, "ts": 7429564831273.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831276.031, "dur": 0.149,
    "args": {
      "External id": 1743,      "cid": 370, "correlation": 613
    }
  },
  {
    "ph": "f", "id": 613, "pid": 537, "tid": 537, "ts": 7429564831276.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831278.056, "dur": 4.558,
    "args": {
      "External id": 1743,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 614,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 614, "pid": 537, "tid": 537, "ts": 7429564831278.056,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831295.015, "dur": 5.183,
    "args": {
      "External id": 1746,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 615,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 615, "pid": 537, "tid": 537, "ts": 7429564831295.015,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831318.112, "dur": 5.419,
    "args": {
      "External id": 1748,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 616,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 616, "pid": 537, "tid": 537, "ts": 7429564831318.112,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831333.946, "dur": 4.178,
    "args": {
      "External id": 1751,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 617,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 617, "pid": 537, "tid": 537, "ts": 7429564831333.946,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831347.337, "dur": 4.740,
    "args": {
      "External id": 1752,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 618,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 618, "pid": 537, "tid": 537, "ts": 7429564831347.337,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831358.272, "dur": 4.585,
    "args": {
      "External id": 1753,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 619,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 619, "pid": 537, "tid": 537, "ts": 7429564831358.272,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831369.394, "dur": 3.688,
    "args": {
      "External id": 1754,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 620,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 620, "pid": 537, "tid": 537, "ts": 7429564831369.394,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831381.982, "dur": 3.407,
    "args": {
      "External id": 1756,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 621,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 621, "pid": 537, "tid": 537, "ts": 7429564831381.982,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564831411.886, "dur": 0.359,
    "args": {
      "External id": 1764,      "cid": 290, "correlation": 622
    }
  },
  {
    "ph": "f", "id": 622, "pid": 537, "tid": 537, "ts": 7429564831411.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564831412.718, "dur": 53.232,
    "args": {
      "External id": 1764,            "bytes": 2097152, "cid": 108, "correlation": 623, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 623, "pid": 537, "tid": 537, "ts": 7429564831412.718,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831492.218, "dur": 0.235,
    "args": {
      "External id": 1764,      "cid": 370, "correlation": 624
    }
  },
  {
    "ph": "f", "id": 624, "pid": 537, "tid": 537, "ts": 7429564831492.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831496.135, "dur": 0.172,
    "args": {
      "External id": 1764,      "cid": 370, "correlation": 625
    }
  },
  {
    "ph": "f", "id": 625, "pid": 537, "tid": 537, "ts": 7429564831496.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831498.832, "dur": 0.168,
    "args": {
      "External id": 1764,      "cid": 370, "correlation": 626
    }
  },
  {
    "ph": "f", "id": 626, "pid": 537, "tid": 537, "ts": 7429564831498.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831507.743, "dur": 5.782,
    "args": {
      "External id": 1764,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 627,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 627, "pid": 537, "tid": 537, "ts": 7429564831507.743,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831514.123, "dur": 4.023,
    "args": {
      "External id": 1764,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 628,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 628, "pid": 537, "tid": 537, "ts": 7429564831514.123,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831539.426, "dur": 5.594,
    "args": {
      "External id": 1767,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 629,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 629, "pid": 537, "tid": 537, "ts": 7429564831539.426,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564831564.530, "dur": 0.308,
    "args": {
      "External id": 1775,      "cid": 290, "correlation": 630
    }
  },
  {
    "ph": "f", "id": 630, "pid": 537, "tid": 537, "ts": 7429564831564.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564831565.214, "dur": 31.898,
    "args": {
      "External id": 1775,            "bytes": 2097152, "cid": 108, "correlation": 631, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 631, "pid": 537, "tid": 537, "ts": 7429564831565.214,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831607.756, "dur": 0.177,
    "args": {
      "External id": 1775,      "cid": 370, "correlation": 632
    }
  },
  {
    "ph": "f", "id": 632, "pid": 537, "tid": 537, "ts": 7429564831607.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831611.253, "dur": 0.141,
    "args": {
      "External id": 1775,      "cid": 370, "correlation": 633
    }
  },
  {
    "ph": "f", "id": 633, "pid": 537, "tid": 537, "ts": 7429564831611.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831614.046, "dur": 0.184,
    "args": {
      "External id": 1775,      "cid": 370, "correlation": 634
    }
  },
  {
    "ph": "f", "id": 634, "pid": 537, "tid": 537, "ts": 7429564831614.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831618.840, "dur": 6.561,
    "args": {
      "External id": 1775,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 635,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 635, "pid": 537, "tid": 537, "ts": 7429564831618.840,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831625.761, "dur": 3.473,
    "args": {
      "External id": 1775,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 636,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 636, "pid": 537, "tid": 537, "ts": 7429564831625.761,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831640.786, "dur": 4.431,
    "args": {
      "External id": 1778,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 637,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 637, "pid": 537, "tid": 537, "ts": 7429564831640.786,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831672.832, "dur": 0.176,
    "args": {
      "External id": 1786,      "cid": 370, "correlation": 638
    }
  },
  {
    "ph": "f", "id": 638, "pid": 537, "tid": 537, "ts": 7429564831672.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831675.949, "dur": 0.140,
    "args": {
      "External id": 1786,      "cid": 370, "correlation": 639
    }
  },
  {
    "ph": "f", "id": 639, "pid": 537, "tid": 537, "ts": 7429564831675.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831678.340, "dur": 0.141,
    "args": {
      "External id": 1786,      "cid": 370, "correlation": 640
    }
  },
  {
    "ph": "f", "id": 640, "pid": 537, "tid": 537, "ts": 7429564831678.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831682.237, "dur": 4.688,
    "args": {
      "External id": 1786,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 641,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 641, "pid": 537, "tid": 537, "ts": 7429564831682.237,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831687.405, "dur": 3.714,
    "args": {
      "External id": 1786,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 642,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 642, "pid": 537, "tid": 537, "ts": 7429564831687.405,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831701.089, "dur": 5.388,
    "args": {
      "External id": 1789,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 643,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 643, "pid": 537, "tid": 537, "ts": 7429564831701.089,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831736.008, "dur": 5.713,
    "args": {
      "External id": 1791,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 644,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 644, "pid": 537, "tid": 537, "ts": 7429564831736.008,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831750.776, "dur": 5.540,
    "args": {
      "External id": 1794,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 645,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 645, "pid": 537, "tid": 537, "ts": 7429564831750.776,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831766.068, "dur": 3.549,
    "args": {
      "External id": 1795,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 646,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 646, "pid": 537, "tid": 537, "ts": 7429564831766.068,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831774.729, "dur": 4.228,
    "args": {
      "External id": 1796,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 647,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 647, "pid": 537, "tid": 537, "ts": 7429564831774.729,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831785.134, "dur": 4.956,
    "args": {
      "External id": 1797,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 648,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 648, "pid": 537, "tid": 537, "ts": 7429564831785.134,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831798.451, "dur": 5.020,
    "args": {
      "External id": 1799,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 649,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 649, "pid": 537, "tid": 537, "ts": 7429564831798.451,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831842.888, "dur": 0.253,
    "args": {
      "External id": 1807,      "cid": 370, "correlation": 650
    }
  },
  {
    "ph": "f", "id": 650, "pid": 537, "tid": 537, "ts": 7429564831842.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831846.562, "dur": 0.186,
    "args": {
      "External id": 1807,      "cid": 370, "correlation": 651
    }
  },
  {
    "ph": "f", "id": 651, "pid": 537, "tid": 537, "ts": 7429564831846.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831848.977, "dur": 0.151,
    "args": {
      "External id": 1807,      "cid": 370, "correlation": 652
    }
  },
  {
    "ph": "f", "id": 652, "pid": 537, "tid": 537, "ts": 7429564831848.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831851.104, "dur": 6.106,
    "args": {
      "External id": 1807,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 653,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 653, "pid": 537, "tid": 537, "ts": 7429564831851.104,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831897.275, "dur": 0.200,
    "args": {
      "External id": 1820,      "cid": 370, "correlation": 654
    }
  },
  {
    "ph": "f", "id": 654, "pid": 537, "tid": 537, "ts": 7429564831897.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831900.134, "dur": 0.158,
    "args": {
      "External id": 1820,      "cid": 370, "correlation": 655
    }
  },
  {
    "ph": "f", "id": 655, "pid": 537, "tid": 537, "ts": 7429564831900.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831902.318, "dur": 0.174,
    "args": {
      "External id": 1820,      "cid": 370, "correlation": 656
    }
  },
  {
    "ph": "f", "id": 656, "pid": 537, "tid": 537, "ts": 7429564831902.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831906.606, "dur": 5.391,
    "args": {
      "External id": 1820,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 657,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 657, "pid": 537, "tid": 537, "ts": 7429564831906.606,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831912.542, "dur": 4.548,
    "args": {
      "External id": 1820,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 658,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 658, "pid": 537, "tid": 537, "ts": 7429564831912.542,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831951.989, "dur": 0.281,
    "args": {
      "External id": 1833,      "cid": 370, "correlation": 659
    }
  },
  {
    "ph": "f", "id": 659, "pid": 537, "tid": 537, "ts": 7429564831951.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831955.166, "dur": 0.141,
    "args": {
      "External id": 1833,      "cid": 370, "correlation": 660
    }
  },
  {
    "ph": "f", "id": 660, "pid": 537, "tid": 537, "ts": 7429564831955.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564831957.088, "dur": 0.153,
    "args": {
      "External id": 1833,      "cid": 370, "correlation": 661
    }
  },
  {
    "ph": "f", "id": 661, "pid": 537, "tid": 537, "ts": 7429564831957.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831960.534, "dur": 4.222,
    "args": {
      "External id": 1833,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 662,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 662, "pid": 537, "tid": 537, "ts": 7429564831960.534,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564831965.043, "dur": 4.897,
    "args": {
      "External id": 1833,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 663,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 663, "pid": 537, "tid": 537, "ts": 7429564831965.043,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564831992.613, "dur": 0.309,
    "args": {
      "External id": 1843,      "cid": 290, "correlation": 664
    }
  },
  {
    "ph": "f", "id": 664, "pid": 537, "tid": 537, "ts": 7429564831992.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564831993.264, "dur": 34.277,
    "args": {
      "External id": 1843,            "bytes": 2097152, "cid": 108, "correlation": 665, "ptr": "0x7fee32880142"
    }
  },
  {
    "ph": "s", "id": 665, "pid": 537, "tid": 537, "ts": 7429564831993.264,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832031.933, "dur": 5.389,
    "args": {
      "External id": 1843,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 666,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 666, "pid": 537, "tid": 537, "ts": 7429564832031.933,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832063.148, "dur": 5.924,
    "args": {
      "External id": 1848,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 667,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 667, "pid": 537, "tid": 537, "ts": 7429564832063.148,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832083.485, "dur": 7.021,
    "args": {
      "External id": 1849,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 668,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 668, "pid": 537, "tid": 537, "ts": 7429564832083.485,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832099.023, "dur": 4.879,
    "args": {
      "External id": 1850,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 669,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 669, "pid": 537, "tid": 537, "ts": 7429564832099.023,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832110.541, "dur": 3.972,
    "args": {
      "External id": 1851,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 670,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 670, "pid": 537, "tid": 537, "ts": 7429564832110.541,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832120.158, "dur": 3.459,
    "args": {
      "External id": 1852,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 671,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 671, "pid": 537, "tid": 537, "ts": 7429564832120.158,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832137.690, "dur": 4.299,
    "args": {
      "External id": 1857,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 672,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 672, "pid": 537, "tid": 537, "ts": 7429564832137.690,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832148.692, "dur": 6.604,
    "args": {
      "External id": 1858,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 673,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 673, "pid": 537, "tid": 537, "ts": 7429564832148.692,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832162.528, "dur": 3.794,
    "args": {
      "External id": 1859,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 674,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 674, "pid": 537, "tid": 537, "ts": 7429564832162.528,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832170.505, "dur": 3.480,
    "args": {
      "External id": 1860,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 675,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 675, "pid": 537, "tid": 537, "ts": 7429564832170.505,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832235.030, "dur": 4.658,
    "args": {
      "External id": 1877,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 676,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 676, "pid": 537, "tid": 537, "ts": 7429564832235.030,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832269.545, "dur": 3.518,
    "args": {
      "External id": 1895,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 677,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 677, "pid": 537, "tid": 537, "ts": 7429564832269.545,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832308.273, "dur": 0.296,
    "args": {
      "External id": 1908,      "cid": 370, "correlation": 678
    }
  },
  {
    "ph": "f", "id": 678, "pid": 537, "tid": 537, "ts": 7429564832308.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832312.981, "dur": 0.162,
    "args": {
      "External id": 1908,      "cid": 370, "correlation": 679
    }
  },
  {
    "ph": "f", "id": 679, "pid": 537, "tid": 537, "ts": 7429564832312.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832315.597, "dur": 0.163,
    "args": {
      "External id": 1908,      "cid": 370, "correlation": 680
    }
  },
  {
    "ph": "f", "id": 680, "pid": 537, "tid": 537, "ts": 7429564832315.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832318.512, "dur": 5.741,
    "args": {
      "External id": 1908,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 681,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 681, "pid": 537, "tid": 537, "ts": 7429564832318.512,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832336.496, "dur": 5.133,
    "args": {
      "External id": 1911,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 682,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 682, "pid": 537, "tid": 537, "ts": 7429564832336.496,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832356.020, "dur": 4.104,
    "args": {
      "External id": 1918,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 683,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 683, "pid": 537, "tid": 537, "ts": 7429564832356.020,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832371.513, "dur": 6.425,
    "args": {
      "External id": 1921,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 684,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 684, "pid": 537, "tid": 537, "ts": 7429564832371.513,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832418.867, "dur": 0.302,
    "args": {
      "External id": 1933,      "cid": 370, "correlation": 685
    }
  },
  {
    "ph": "f", "id": 685, "pid": 537, "tid": 537, "ts": 7429564832418.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832422.709, "dur": 0.148,
    "args": {
      "External id": 1933,      "cid": 370, "correlation": 686
    }
  },
  {
    "ph": "f", "id": 686, "pid": 537, "tid": 537, "ts": 7429564832422.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832425.460, "dur": 0.162,
    "args": {
      "External id": 1933,      "cid": 370, "correlation": 687
    }
  },
  {
    "ph": "f", "id": 687, "pid": 537, "tid": 537, "ts": 7429564832425.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832427.496, "dur": 6.184,
    "args": {
      "External id": 1933,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 688,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 688, "pid": 537, "tid": 537, "ts": 7429564832427.496,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832449.914, "dur": 3.991,
    "args": {
      "External id": 1942,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 689,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 689, "pid": 537, "tid": 537, "ts": 7429564832449.914,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832494.907, "dur": 0.297,
    "args": {
      "External id": 1952,      "cid": 370, "correlation": 690
    }
  },
  {
    "ph": "f", "id": 690, "pid": 537, "tid": 537, "ts": 7429564832494.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832497.961, "dur": 0.161,
    "args": {
      "External id": 1952,      "cid": 370, "correlation": 691
    }
  },
  {
    "ph": "f", "id": 691, "pid": 537, "tid": 537, "ts": 7429564832497.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832500.340, "dur": 0.178,
    "args": {
      "External id": 1952,      "cid": 370, "correlation": 692
    }
  },
  {
    "ph": "f", "id": 692, "pid": 537, "tid": 537, "ts": 7429564832500.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832502.225, "dur": 5.306,
    "args": {
      "External id": 1952,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 693,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 693, "pid": 537, "tid": 537, "ts": 7429564832502.225,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832519.475, "dur": 5.500,
    "args": {
      "External id": 1955,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 694,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 694, "pid": 537, "tid": 537, "ts": 7429564832519.475,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832543.038, "dur": 4.197,
    "args": {
      "External id": 1957,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 695,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 695, "pid": 537, "tid": 537, "ts": 7429564832543.038,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832557.413, "dur": 6.469,
    "args": {
      "External id": 1960,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 696,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 696, "pid": 537, "tid": 537, "ts": 7429564832557.413,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832572.371, "dur": 3.947,
    "args": {
      "External id": 1961,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 697,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 697, "pid": 537, "tid": 537, "ts": 7429564832572.371,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832583.383, "dur": 5.855,
    "args": {
      "External id": 1962,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 698,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 698, "pid": 537, "tid": 537, "ts": 7429564832583.383,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832595.199, "dur": 4.200,
    "args": {
      "External id": 1963,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 699,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 699, "pid": 537, "tid": 537, "ts": 7429564832595.199,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832608.030, "dur": 5.483,
    "args": {
      "External id": 1965,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 700,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 700, "pid": 537, "tid": 537, "ts": 7429564832608.030,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832649.633, "dur": 0.183,
    "args": {
      "External id": 1973,      "cid": 370, "correlation": 701
    }
  },
  {
    "ph": "f", "id": 701, "pid": 537, "tid": 537, "ts": 7429564832649.633,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832652.846, "dur": 0.140,
    "args": {
      "External id": 1973,      "cid": 370, "correlation": 702
    }
  },
  {
    "ph": "f", "id": 702, "pid": 537, "tid": 537, "ts": 7429564832652.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832654.960, "dur": 0.157,
    "args": {
      "External id": 1973,      "cid": 370, "correlation": 703
    }
  },
  {
    "ph": "f", "id": 703, "pid": 537, "tid": 537, "ts": 7429564832654.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832662.207, "dur": 5.528,
    "args": {
      "External id": 1973,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 704,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 704, "pid": 537, "tid": 537, "ts": 7429564832662.207,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832668.081, "dur": 5.530,
    "args": {
      "External id": 1973,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 705,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 705, "pid": 537, "tid": 537, "ts": 7429564832668.081,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564832690.578, "dur": 0.337,
    "args": {
      "External id": 1976,      "cid": 290, "correlation": 706
    }
  },
  {
    "ph": "f", "id": 706, "pid": 537, "tid": 537, "ts": 7429564832690.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564832691.281, "dur": 47.800,
    "args": {
      "External id": 1976,            "bytes": 2097152, "cid": 108, "correlation": 707, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 707, "pid": 537, "tid": 537, "ts": 7429564832691.281,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832742.734, "dur": 4.667,
    "args": {
      "External id": 1976,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 708,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 708, "pid": 537, "tid": 537, "ts": 7429564832742.734,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832778.131, "dur": 0.211,
    "args": {
      "External id": 1984,      "cid": 370, "correlation": 709
    }
  },
  {
    "ph": "f", "id": 709, "pid": 537, "tid": 537, "ts": 7429564832778.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832781.389, "dur": 0.151,
    "args": {
      "External id": 1984,      "cid": 370, "correlation": 710
    }
  },
  {
    "ph": "f", "id": 710, "pid": 537, "tid": 537, "ts": 7429564832781.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832783.629, "dur": 0.162,
    "args": {
      "External id": 1984,      "cid": 370, "correlation": 711
    }
  },
  {
    "ph": "f", "id": 711, "pid": 537, "tid": 537, "ts": 7429564832783.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832788.725, "dur": 5.901,
    "args": {
      "External id": 1984,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 712,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 712, "pid": 537, "tid": 537, "ts": 7429564832788.725,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832794.957, "dur": 4.418,
    "args": {
      "External id": 1984,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 713,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 713, "pid": 537, "tid": 537, "ts": 7429564832794.957,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564832808.306, "dur": 0.240,
    "args": {
      "External id": 1987,      "cid": 290, "correlation": 714
    }
  },
  {
    "ph": "f", "id": 714, "pid": 537, "tid": 537, "ts": 7429564832808.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564832808.784, "dur": 27.307,
    "args": {
      "External id": 1987,            "bytes": 2097152, "cid": 108, "correlation": 715, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 715, "pid": 537, "tid": 537, "ts": 7429564832808.784,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832838.649, "dur": 5.031,
    "args": {
      "External id": 1987,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 716,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 716, "pid": 537, "tid": 537, "ts": 7429564832838.649,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832870.835, "dur": 0.210,
    "args": {
      "External id": 1995,      "cid": 370, "correlation": 717
    }
  },
  {
    "ph": "f", "id": 717, "pid": 537, "tid": 537, "ts": 7429564832870.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832874.309, "dur": 0.158,
    "args": {
      "External id": 1995,      "cid": 370, "correlation": 718
    }
  },
  {
    "ph": "f", "id": 718, "pid": 537, "tid": 537, "ts": 7429564832874.309,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564832876.628, "dur": 0.146,
    "args": {
      "External id": 1995,      "cid": 370, "correlation": 719
    }
  },
  {
    "ph": "f", "id": 719, "pid": 537, "tid": 537, "ts": 7429564832876.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832880.824, "dur": 4.874,
    "args": {
      "External id": 1995,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 720,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 720, "pid": 537, "tid": 537, "ts": 7429564832880.824,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832885.976, "dur": 4.796,
    "args": {
      "External id": 1995,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 721,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 721, "pid": 537, "tid": 537, "ts": 7429564832885.976,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832901.565, "dur": 4.121,
    "args": {
      "External id": 1998,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 722,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 722, "pid": 537, "tid": 537, "ts": 7429564832901.565,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832932.818, "dur": 4.067,
    "args": {
      "External id": 2000,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 723,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 723, "pid": 537, "tid": 537, "ts": 7429564832932.818,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832947.208, "dur": 4.182,
    "args": {
      "External id": 2003,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 724,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 724, "pid": 537, "tid": 537, "ts": 7429564832947.208,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832960.363, "dur": 4.713,
    "args": {
      "External id": 2004,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 725,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 725, "pid": 537, "tid": 537, "ts": 7429564832960.363,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832970.333, "dur": 4.068,
    "args": {
      "External id": 2005,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 726,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 726, "pid": 537, "tid": 537, "ts": 7429564832970.333,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832980.573, "dur": 4.149,
    "args": {
      "External id": 2006,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 727,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 727, "pid": 537, "tid": 537, "ts": 7429564832980.573,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564832993.104, "dur": 3.177,
    "args": {
      "External id": 2008,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 728,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 728, "pid": 537, "tid": 537, "ts": 7429564832993.104,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833036.947, "dur": 0.287,
    "args": {
      "External id": 2016,      "cid": 370, "correlation": 729
    }
  },
  {
    "ph": "f", "id": 729, "pid": 537, "tid": 537, "ts": 7429564833036.947,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833041.181, "dur": 0.177,
    "args": {
      "External id": 2016,      "cid": 370, "correlation": 730
    }
  },
  {
    "ph": "f", "id": 730, "pid": 537, "tid": 537, "ts": 7429564833041.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833043.674, "dur": 0.189,
    "args": {
      "External id": 2016,      "cid": 370, "correlation": 731
    }
  },
  {
    "ph": "f", "id": 731, "pid": 537, "tid": 537, "ts": 7429564833043.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833046.479, "dur": 4.587,
    "args": {
      "External id": 2016,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 732,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 732, "pid": 537, "tid": 537, "ts": 7429564833046.479,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833244.379, "dur": 0.359,
    "args": {
      "External id": 2029,      "cid": 370, "correlation": 733
    }
  },
  {
    "ph": "f", "id": 733, "pid": 537, "tid": 537, "ts": 7429564833244.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833248.688, "dur": 0.154,
    "args": {
      "External id": 2029,      "cid": 370, "correlation": 734
    }
  },
  {
    "ph": "f", "id": 734, "pid": 537, "tid": 537, "ts": 7429564833248.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833250.748, "dur": 0.314,
    "args": {
      "External id": 2029,      "cid": 370, "correlation": 735
    }
  },
  {
    "ph": "f", "id": 735, "pid": 537, "tid": 537, "ts": 7429564833250.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833256.902, "dur": 5.964,
    "args": {
      "External id": 2029,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 736,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 736, "pid": 537, "tid": 537, "ts": 7429564833256.902,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833263.369, "dur": 4.928,
    "args": {
      "External id": 2029,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 737,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 737, "pid": 537, "tid": 537, "ts": 7429564833263.369,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833310.351, "dur": 0.237,
    "args": {
      "External id": 2042,      "cid": 370, "correlation": 738
    }
  },
  {
    "ph": "f", "id": 738, "pid": 537, "tid": 537, "ts": 7429564833310.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833313.301, "dur": 0.142,
    "args": {
      "External id": 2042,      "cid": 370, "correlation": 739
    }
  },
  {
    "ph": "f", "id": 739, "pid": 537, "tid": 537, "ts": 7429564833313.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833315.802, "dur": 0.172,
    "args": {
      "External id": 2042,      "cid": 370, "correlation": 740
    }
  },
  {
    "ph": "f", "id": 740, "pid": 537, "tid": 537, "ts": 7429564833315.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833319.640, "dur": 4.710,
    "args": {
      "External id": 2042,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 741,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 741, "pid": 537, "tid": 537, "ts": 7429564833319.640,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833324.795, "dur": 4.183,
    "args": {
      "External id": 2042,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 742,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 742, "pid": 537, "tid": 537, "ts": 7429564833324.795,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833385.325, "dur": 4.165,
    "args": {
      "External id": 2052,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 743,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 743, "pid": 537, "tid": 537, "ts": 7429564833385.325,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833413.703, "dur": 6.533,
    "args": {
      "External id": 2057,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 744,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 744, "pid": 537, "tid": 537, "ts": 7429564833413.703,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833436.501, "dur": 7.317,
    "args": {
      "External id": 2058,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 745,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 745, "pid": 537, "tid": 537, "ts": 7429564833436.501,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833451.553, "dur": 4.789,
    "args": {
      "External id": 2059,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 746,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 746, "pid": 537, "tid": 537, "ts": 7429564833451.553,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833462.423, "dur": 4.938,
    "args": {
      "External id": 2060,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 747,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 747, "pid": 537, "tid": 537, "ts": 7429564833462.423,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833473.978, "dur": 3.447,
    "args": {
      "External id": 2061,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 748,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 748, "pid": 537, "tid": 537, "ts": 7429564833473.978,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833489.631, "dur": 4.660,
    "args": {
      "External id": 2066,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 749,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 749, "pid": 537, "tid": 537, "ts": 7429564833489.631,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833500.488, "dur": 5.412,
    "args": {
      "External id": 2067,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 750,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 750, "pid": 537, "tid": 537, "ts": 7429564833500.488,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833511.679, "dur": 4.805,
    "args": {
      "External id": 2068,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 751,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 751, "pid": 537, "tid": 537, "ts": 7429564833511.679,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833521.726, "dur": 4.735,
    "args": {
      "External id": 2069,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 752,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 752, "pid": 537, "tid": 537, "ts": 7429564833521.726,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833589.463, "dur": 5.026,
    "args": {
      "External id": 2086,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 753,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 753, "pid": 537, "tid": 537, "ts": 7429564833589.463,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833624.071, "dur": 5.142,
    "args": {
      "External id": 2104,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 754,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 754, "pid": 537, "tid": 537, "ts": 7429564833624.071,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833665.554, "dur": 0.292,
    "args": {
      "External id": 2117,      "cid": 370, "correlation": 755
    }
  },
  {
    "ph": "f", "id": 755, "pid": 537, "tid": 537, "ts": 7429564833665.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833670.301, "dur": 0.146,
    "args": {
      "External id": 2117,      "cid": 370, "correlation": 756
    }
  },
  {
    "ph": "f", "id": 756, "pid": 537, "tid": 537, "ts": 7429564833670.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833672.867, "dur": 0.191,
    "args": {
      "External id": 2117,      "cid": 370, "correlation": 757
    }
  },
  {
    "ph": "f", "id": 757, "pid": 537, "tid": 537, "ts": 7429564833672.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833675.684, "dur": 4.572,
    "args": {
      "External id": 2117,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 758,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 758, "pid": 537, "tid": 537, "ts": 7429564833675.684,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833692.394, "dur": 5.169,
    "args": {
      "External id": 2120,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 759,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 759, "pid": 537, "tid": 537, "ts": 7429564833692.394,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833711.495, "dur": 3.738,
    "args": {
      "External id": 2127,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 760,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 760, "pid": 537, "tid": 537, "ts": 7429564833711.495,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833727.061, "dur": 7.200,
    "args": {
      "External id": 2130,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 761,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 761, "pid": 537, "tid": 537, "ts": 7429564833727.061,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833775.383, "dur": 0.174,
    "args": {
      "External id": 2142,      "cid": 370, "correlation": 762
    }
  },
  {
    "ph": "f", "id": 762, "pid": 537, "tid": 537, "ts": 7429564833775.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833778.951, "dur": 0.153,
    "args": {
      "External id": 2142,      "cid": 370, "correlation": 763
    }
  },
  {
    "ph": "f", "id": 763, "pid": 537, "tid": 537, "ts": 7429564833778.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833782.112, "dur": 0.171,
    "args": {
      "External id": 2142,      "cid": 370, "correlation": 764
    }
  },
  {
    "ph": "f", "id": 764, "pid": 537, "tid": 537, "ts": 7429564833782.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833784.109, "dur": 5.223,
    "args": {
      "External id": 2142,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 765,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 765, "pid": 537, "tid": 537, "ts": 7429564833784.109,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833806.460, "dur": 3.702,
    "args": {
      "External id": 2151,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 766,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 766, "pid": 537, "tid": 537, "ts": 7429564833806.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833852.574, "dur": 0.191,
    "args": {
      "External id": 2161,      "cid": 370, "correlation": 767
    }
  },
  {
    "ph": "f", "id": 767, "pid": 537, "tid": 537, "ts": 7429564833852.574,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833855.705, "dur": 0.172,
    "args": {
      "External id": 2161,      "cid": 370, "correlation": 768
    }
  },
  {
    "ph": "f", "id": 768, "pid": 537, "tid": 537, "ts": 7429564833855.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564833857.876, "dur": 0.150,
    "args": {
      "External id": 2161,      "cid": 370, "correlation": 769
    }
  },
  {
    "ph": "f", "id": 769, "pid": 537, "tid": 537, "ts": 7429564833857.876,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833859.817, "dur": 4.391,
    "args": {
      "External id": 2161,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 770,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 770, "pid": 537, "tid": 537, "ts": 7429564833859.817,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564833877.051, "dur": 0.343,
    "args": {
      "External id": 2164,      "cid": 290, "correlation": 771
    }
  },
  {
    "ph": "f", "id": 771, "pid": 537, "tid": 537, "ts": 7429564833877.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564833878.070, "dur": 63.025,
    "args": {
      "External id": 2164,            "bytes": 2097152, "cid": 108, "correlation": 772, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 772, "pid": 537, "tid": 537, "ts": 7429564833878.070,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833943.930, "dur": 4.553,
    "args": {
      "External id": 2164,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 773,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 773, "pid": 537, "tid": 537, "ts": 7429564833943.930,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564833984.935, "dur": 6.213,
    "args": {
      "External id": 2166,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 774,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 774, "pid": 537, "tid": 537, "ts": 7429564833984.935,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834002.408, "dur": 6.705,
    "args": {
      "External id": 2169,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 775,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 775, "pid": 537, "tid": 537, "ts": 7429564834002.408,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834017.811, "dur": 4.223,
    "args": {
      "External id": 2170,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 776,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 776, "pid": 537, "tid": 537, "ts": 7429564834017.811,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834028.613, "dur": 4.820,
    "args": {
      "External id": 2171,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 777,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 777, "pid": 537, "tid": 537, "ts": 7429564834028.613,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834040.491, "dur": 4.802,
    "args": {
      "External id": 2172,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 778,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 778, "pid": 537, "tid": 537, "ts": 7429564834040.491,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834052.610, "dur": 3.694,
    "args": {
      "External id": 2174,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 779,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 779, "pid": 537, "tid": 537, "ts": 7429564834052.610,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834123.867, "dur": 0.446,
    "args": {
      "External id": 2182,      "cid": 370, "correlation": 780
    }
  },
  {
    "ph": "f", "id": 780, "pid": 537, "tid": 537, "ts": 7429564834123.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834128.900, "dur": 0.185,
    "args": {
      "External id": 2182,      "cid": 370, "correlation": 781
    }
  },
  {
    "ph": "f", "id": 781, "pid": 537, "tid": 537, "ts": 7429564834128.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834131.964, "dur": 0.335,
    "args": {
      "External id": 2182,      "cid": 370, "correlation": 782
    }
  },
  {
    "ph": "f", "id": 782, "pid": 537, "tid": 537, "ts": 7429564834131.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834143.076, "dur": 6.494,
    "args": {
      "External id": 2182,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 783,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 783, "pid": 537, "tid": 537, "ts": 7429564834143.076,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834150.033, "dur": 4.964,
    "args": {
      "External id": 2182,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 784,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 784, "pid": 537, "tid": 537, "ts": 7429564834150.033,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834175.976, "dur": 4.902,
    "args": {
      "External id": 2185,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 785,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 785, "pid": 537, "tid": 537, "ts": 7429564834175.976,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564834200.896, "dur": 0.378,
    "args": {
      "External id": 2193,      "cid": 290, "correlation": 786
    }
  },
  {
    "ph": "f", "id": 786, "pid": 537, "tid": 537, "ts": 7429564834200.896,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564834201.746, "dur": 44.362,
    "args": {
      "External id": 2193,            "bytes": 2097152, "cid": 108, "correlation": 787, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 787, "pid": 537, "tid": 537, "ts": 7429564834201.746,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834258.878, "dur": 0.210,
    "args": {
      "External id": 2193,      "cid": 370, "correlation": 788
    }
  },
  {
    "ph": "f", "id": 788, "pid": 537, "tid": 537, "ts": 7429564834258.878,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834262.564, "dur": 0.142,
    "args": {
      "External id": 2193,      "cid": 370, "correlation": 789
    }
  },
  {
    "ph": "f", "id": 789, "pid": 537, "tid": 537, "ts": 7429564834262.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834265.220, "dur": 0.162,
    "args": {
      "External id": 2193,      "cid": 370, "correlation": 790
    }
  },
  {
    "ph": "f", "id": 790, "pid": 537, "tid": 537, "ts": 7429564834265.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834270.596, "dur": 6.529,
    "args": {
      "External id": 2193,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 791,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 791, "pid": 537, "tid": 537, "ts": 7429564834270.596,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834277.462, "dur": 3.761,
    "args": {
      "External id": 2193,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 792,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 792, "pid": 537, "tid": 537, "ts": 7429564834277.462,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834292.860, "dur": 4.586,
    "args": {
      "External id": 2196,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 793,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 793, "pid": 537, "tid": 537, "ts": 7429564834292.860,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834327.068, "dur": 0.176,
    "args": {
      "External id": 2204,      "cid": 370, "correlation": 794
    }
  },
  {
    "ph": "f", "id": 794, "pid": 537, "tid": 537, "ts": 7429564834327.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834330.084, "dur": 0.139,
    "args": {
      "External id": 2204,      "cid": 370, "correlation": 795
    }
  },
  {
    "ph": "f", "id": 795, "pid": 537, "tid": 537, "ts": 7429564834330.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834332.112, "dur": 0.159,
    "args": {
      "External id": 2204,      "cid": 370, "correlation": 796
    }
  },
  {
    "ph": "f", "id": 796, "pid": 537, "tid": 537, "ts": 7429564834332.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834335.880, "dur": 4.837,
    "args": {
      "External id": 2204,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 797,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 797, "pid": 537, "tid": 537, "ts": 7429564834335.880,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834341.179, "dur": 4.850,
    "args": {
      "External id": 2204,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 798,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 798, "pid": 537, "tid": 537, "ts": 7429564834341.179,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834356.712, "dur": 4.444,
    "args": {
      "External id": 2207,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 799,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 799, "pid": 537, "tid": 537, "ts": 7429564834356.712,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834391.019, "dur": 4.270,
    "args": {
      "External id": 2209,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 800,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 800, "pid": 537, "tid": 537, "ts": 7429564834391.019,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834405.737, "dur": 4.879,
    "args": {
      "External id": 2212,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 801,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 801, "pid": 537, "tid": 537, "ts": 7429564834405.737,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834419.154, "dur": 4.767,
    "args": {
      "External id": 2213,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 802,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 802, "pid": 537, "tid": 537, "ts": 7429564834419.154,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834429.406, "dur": 5.139,
    "args": {
      "External id": 2214,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 803,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 803, "pid": 537, "tid": 537, "ts": 7429564834429.406,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834441.701, "dur": 4.842,
    "args": {
      "External id": 2215,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 804,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 804, "pid": 537, "tid": 537, "ts": 7429564834441.701,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564834453.829, "dur": 0.233,
    "args": {
      "External id": 2217,      "cid": 290, "correlation": 805
    }
  },
  {
    "ph": "f", "id": 805, "pid": 537, "tid": 537, "ts": 7429564834453.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564834454.465, "dur": 34.856,
    "args": {
      "External id": 2217,            "bytes": 2097152, "cid": 108, "correlation": 806, "ptr": "0x100000002"
    }
  },
  {
    "ph": "s", "id": 806, "pid": 537, "tid": 537, "ts": 7429564834454.465,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834491.306, "dur": 4.281,
    "args": {
      "External id": 2217,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 807,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 807, "pid": 537, "tid": 537, "ts": 7429564834491.306,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834542.565, "dur": 0.315,
    "args": {
      "External id": 2225,      "cid": 370, "correlation": 808
    }
  },
  {
    "ph": "f", "id": 808, "pid": 537, "tid": 537, "ts": 7429564834542.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834547.327, "dur": 0.145,
    "args": {
      "External id": 2225,      "cid": 370, "correlation": 809
    }
  },
  {
    "ph": "f", "id": 809, "pid": 537, "tid": 537, "ts": 7429564834547.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834550.204, "dur": 0.369,
    "args": {
      "External id": 2225,      "cid": 370, "correlation": 810
    }
  },
  {
    "ph": "f", "id": 810, "pid": 537, "tid": 537, "ts": 7429564834550.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834553.409, "dur": 6.286,
    "args": {
      "External id": 2225,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 811,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 811, "pid": 537, "tid": 537, "ts": 7429564834553.409,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834601.215, "dur": 0.185,
    "args": {
      "External id": 2238,      "cid": 370, "correlation": 812
    }
  },
  {
    "ph": "f", "id": 812, "pid": 537, "tid": 537, "ts": 7429564834601.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834604.935, "dur": 0.142,
    "args": {
      "External id": 2238,      "cid": 370, "correlation": 813
    }
  },
  {
    "ph": "f", "id": 813, "pid": 537, "tid": 537, "ts": 7429564834604.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834607.520, "dur": 0.151,
    "args": {
      "External id": 2238,      "cid": 370, "correlation": 814
    }
  },
  {
    "ph": "f", "id": 814, "pid": 537, "tid": 537, "ts": 7429564834607.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834612.757, "dur": 5.998,
    "args": {
      "External id": 2238,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 815,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 815, "pid": 537, "tid": 537, "ts": 7429564834612.757,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834619.168, "dur": 4.109,
    "args": {
      "External id": 2238,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 816,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 816, "pid": 537, "tid": 537, "ts": 7429564834619.168,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834657.163, "dur": 0.245,
    "args": {
      "External id": 2251,      "cid": 370, "correlation": 817
    }
  },
  {
    "ph": "f", "id": 817, "pid": 537, "tid": 537, "ts": 7429564834657.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834660.211, "dur": 0.150,
    "args": {
      "External id": 2251,      "cid": 370, "correlation": 818
    }
  },
  {
    "ph": "f", "id": 818, "pid": 537, "tid": 537, "ts": 7429564834660.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834662.218, "dur": 0.147,
    "args": {
      "External id": 2251,      "cid": 370, "correlation": 819
    }
  },
  {
    "ph": "f", "id": 819, "pid": 537, "tid": 537, "ts": 7429564834662.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834666.142, "dur": 4.817,
    "args": {
      "External id": 2251,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 820,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 820, "pid": 537, "tid": 537, "ts": 7429564834666.142,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834671.251, "dur": 4.497,
    "args": {
      "External id": 2251,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 821,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 821, "pid": 537, "tid": 537, "ts": 7429564834671.251,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834699.624, "dur": 4.342,
    "args": {
      "External id": 2261,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 822,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 822, "pid": 537, "tid": 537, "ts": 7429564834699.624,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834728.050, "dur": 4.691,
    "args": {
      "External id": 2266,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 823,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 823, "pid": 537, "tid": 537, "ts": 7429564834728.050,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834748.183, "dur": 7.149,
    "args": {
      "External id": 2267,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 824,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 824, "pid": 537, "tid": 537, "ts": 7429564834748.183,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834763.730, "dur": 4.158,
    "args": {
      "External id": 2268,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 825,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 825, "pid": 537, "tid": 537, "ts": 7429564834763.730,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834773.291, "dur": 5.205,
    "args": {
      "External id": 2269,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 826,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 826, "pid": 537, "tid": 537, "ts": 7429564834773.291,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834785.362, "dur": 4.853,
    "args": {
      "External id": 2270,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 827,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 827, "pid": 537, "tid": 537, "ts": 7429564834785.362,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834804.396, "dur": 3.512,
    "args": {
      "External id": 2275,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 828,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 828, "pid": 537, "tid": 537, "ts": 7429564834804.396,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834814.144, "dur": 4.822,
    "args": {
      "External id": 2276,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 829,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 829, "pid": 537, "tid": 537, "ts": 7429564834814.144,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834825.485, "dur": 3.850,
    "args": {
      "External id": 2277,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 830,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 830, "pid": 537, "tid": 537, "ts": 7429564834825.485,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834833.930, "dur": 3.360,
    "args": {
      "External id": 2278,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 831,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 831, "pid": 537, "tid": 537, "ts": 7429564834833.930,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834899.905, "dur": 4.951,
    "args": {
      "External id": 2295,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 832,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 832, "pid": 537, "tid": 537, "ts": 7429564834899.905,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834933.689, "dur": 5.696,
    "args": {
      "External id": 2313,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 833,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 833, "pid": 537, "tid": 537, "ts": 7429564834933.689,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834979.481, "dur": 0.251,
    "args": {
      "External id": 2326,      "cid": 370, "correlation": 834
    }
  },
  {
    "ph": "f", "id": 834, "pid": 537, "tid": 537, "ts": 7429564834979.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834983.484, "dur": 0.166,
    "args": {
      "External id": 2326,      "cid": 370, "correlation": 835
    }
  },
  {
    "ph": "f", "id": 835, "pid": 537, "tid": 537, "ts": 7429564834983.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564834986.049, "dur": 0.147,
    "args": {
      "External id": 2326,      "cid": 370, "correlation": 836
    }
  },
  {
    "ph": "f", "id": 836, "pid": 537, "tid": 537, "ts": 7429564834986.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564834988.544, "dur": 5.559,
    "args": {
      "External id": 2326,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 837,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 837, "pid": 537, "tid": 537, "ts": 7429564834988.544,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835006.446, "dur": 4.708,
    "args": {
      "External id": 2329,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 838,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 838, "pid": 537, "tid": 537, "ts": 7429564835006.446,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835025.336, "dur": 3.845,
    "args": {
      "External id": 2336,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 839,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 839, "pid": 537, "tid": 537, "ts": 7429564835025.336,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835041.314, "dur": 5.552,
    "args": {
      "External id": 2339,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 840,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 840, "pid": 537, "tid": 537, "ts": 7429564835041.314,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835088.183, "dur": 0.203,
    "args": {
      "External id": 2351,      "cid": 370, "correlation": 841
    }
  },
  {
    "ph": "f", "id": 841, "pid": 537, "tid": 537, "ts": 7429564835088.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835092.027, "dur": 0.157,
    "args": {
      "External id": 2351,      "cid": 370, "correlation": 842
    }
  },
  {
    "ph": "f", "id": 842, "pid": 537, "tid": 537, "ts": 7429564835092.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835094.592, "dur": 0.148,
    "args": {
      "External id": 2351,      "cid": 370, "correlation": 843
    }
  },
  {
    "ph": "f", "id": 843, "pid": 537, "tid": 537, "ts": 7429564835094.592,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835096.683, "dur": 4.357,
    "args": {
      "External id": 2351,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 844,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 844, "pid": 537, "tid": 537, "ts": 7429564835096.683,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835118.065, "dur": 4.091,
    "args": {
      "External id": 2360,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 845,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 845, "pid": 537, "tid": 537, "ts": 7429564835118.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835165.128, "dur": 0.278,
    "args": {
      "External id": 2370,      "cid": 370, "correlation": 846
    }
  },
  {
    "ph": "f", "id": 846, "pid": 537, "tid": 537, "ts": 7429564835165.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835168.192, "dur": 0.144,
    "args": {
      "External id": 2370,      "cid": 370, "correlation": 847
    }
  },
  {
    "ph": "f", "id": 847, "pid": 537, "tid": 537, "ts": 7429564835168.192,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835170.279, "dur": 0.163,
    "args": {
      "External id": 2370,      "cid": 370, "correlation": 848
    }
  },
  {
    "ph": "f", "id": 848, "pid": 537, "tid": 537, "ts": 7429564835170.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835172.223, "dur": 4.434,
    "args": {
      "External id": 2370,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 849,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 849, "pid": 537, "tid": 537, "ts": 7429564835172.223,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835189.475, "dur": 5.164,
    "args": {
      "External id": 2373,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 850,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 850, "pid": 537, "tid": 537, "ts": 7429564835189.475,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835211.819, "dur": 5.659,
    "args": {
      "External id": 2375,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 851,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 851, "pid": 537, "tid": 537, "ts": 7429564835211.819,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835227.013, "dur": 4.941,
    "args": {
      "External id": 2378,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 852,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 852, "pid": 537, "tid": 537, "ts": 7429564835227.013,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835240.992, "dur": 3.659,
    "args": {
      "External id": 2379,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 853,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 853, "pid": 537, "tid": 537, "ts": 7429564835240.992,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835251.540, "dur": 5.597,
    "args": {
      "External id": 2380,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 854,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 854, "pid": 537, "tid": 537, "ts": 7429564835251.540,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835263.478, "dur": 3.947,
    "args": {
      "External id": 2381,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 855,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 855, "pid": 537, "tid": 537, "ts": 7429564835263.478,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835276.172, "dur": 3.832,
    "args": {
      "External id": 2383,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 856,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 856, "pid": 537, "tid": 537, "ts": 7429564835276.172,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564835305.747, "dur": 0.322,
    "args": {
      "External id": 2391,      "cid": 290, "correlation": 857
    }
  },
  {
    "ph": "f", "id": 857, "pid": 537, "tid": 537, "ts": 7429564835305.747,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564835306.529, "dur": 50.388,
    "args": {
      "External id": 2391,            "bytes": 2097152, "cid": 108, "correlation": 858, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 858, "pid": 537, "tid": 537, "ts": 7429564835306.529,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835372.213, "dur": 0.256,
    "args": {
      "External id": 2391,      "cid": 370, "correlation": 859
    }
  },
  {
    "ph": "f", "id": 859, "pid": 537, "tid": 537, "ts": 7429564835372.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835376.204, "dur": 0.165,
    "args": {
      "External id": 2391,      "cid": 370, "correlation": 860
    }
  },
  {
    "ph": "f", "id": 860, "pid": 537, "tid": 537, "ts": 7429564835376.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835378.847, "dur": 0.173,
    "args": {
      "External id": 2391,      "cid": 370, "correlation": 861
    }
  },
  {
    "ph": "f", "id": 861, "pid": 537, "tid": 537, "ts": 7429564835378.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835387.254, "dur": 6.595,
    "args": {
      "External id": 2391,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 862,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 862, "pid": 537, "tid": 537, "ts": 7429564835387.254,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835394.194, "dur": 3.666,
    "args": {
      "External id": 2391,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 863,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 863, "pid": 537, "tid": 537, "ts": 7429564835394.194,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835417.650, "dur": 4.624,
    "args": {
      "External id": 2394,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 864,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 864, "pid": 537, "tid": 537, "ts": 7429564835417.650,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564835442.047, "dur": 0.217,
    "args": {
      "External id": 2402,      "cid": 290, "correlation": 865
    }
  },
  {
    "ph": "f", "id": 865, "pid": 537, "tid": 537, "ts": 7429564835442.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564835442.652, "dur": 28.993,
    "args": {
      "External id": 2402,            "bytes": 2097152, "cid": 108, "correlation": 866, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 866, "pid": 537, "tid": 537, "ts": 7429564835442.652,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835484.359, "dur": 0.178,
    "args": {
      "External id": 2402,      "cid": 370, "correlation": 867
    }
  },
  {
    "ph": "f", "id": 867, "pid": 537, "tid": 537, "ts": 7429564835484.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835487.772, "dur": 0.150,
    "args": {
      "External id": 2402,      "cid": 370, "correlation": 868
    }
  },
  {
    "ph": "f", "id": 868, "pid": 537, "tid": 537, "ts": 7429564835487.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835490.213, "dur": 0.144,
    "args": {
      "External id": 2402,      "cid": 370, "correlation": 869
    }
  },
  {
    "ph": "f", "id": 869, "pid": 537, "tid": 537, "ts": 7429564835490.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835495.037, "dur": 6.044,
    "args": {
      "External id": 2402,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 870,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 870, "pid": 537, "tid": 537, "ts": 7429564835495.037,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835501.427, "dur": 3.815,
    "args": {
      "External id": 2402,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 871,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 871, "pid": 537, "tid": 537, "ts": 7429564835501.427,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835516.315, "dur": 4.110,
    "args": {
      "External id": 2405,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 872,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 872, "pid": 537, "tid": 537, "ts": 7429564835516.315,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835544.824, "dur": 0.233,
    "args": {
      "External id": 2413,      "cid": 370, "correlation": 873
    }
  },
  {
    "ph": "f", "id": 873, "pid": 537, "tid": 537, "ts": 7429564835544.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835548.682, "dur": 0.141,
    "args": {
      "External id": 2413,      "cid": 370, "correlation": 874
    }
  },
  {
    "ph": "f", "id": 874, "pid": 537, "tid": 537, "ts": 7429564835548.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835550.565, "dur": 0.146,
    "args": {
      "External id": 2413,      "cid": 370, "correlation": 875
    }
  },
  {
    "ph": "f", "id": 875, "pid": 537, "tid": 537, "ts": 7429564835550.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835554.376, "dur": 4.453,
    "args": {
      "External id": 2413,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 876,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 876, "pid": 537, "tid": 537, "ts": 7429564835554.376,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835559.139, "dur": 3.300,
    "args": {
      "External id": 2413,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 877,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 877, "pid": 537, "tid": 537, "ts": 7429564835559.139,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835573.026, "dur": 5.017,
    "args": {
      "External id": 2416,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 878,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 878, "pid": 537, "tid": 537, "ts": 7429564835573.026,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835605.449, "dur": 4.995,
    "args": {
      "External id": 2418,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 879,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 879, "pid": 537, "tid": 537, "ts": 7429564835605.449,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835619.911, "dur": 5.045,
    "args": {
      "External id": 2421,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 880,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 880, "pid": 537, "tid": 537, "ts": 7429564835619.911,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835632.963, "dur": 3.613,
    "args": {
      "External id": 2422,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 881,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 881, "pid": 537, "tid": 537, "ts": 7429564835632.963,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835641.691, "dur": 5.575,
    "args": {
      "External id": 2423,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 882,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 882, "pid": 537, "tid": 537, "ts": 7429564835641.691,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835653.662, "dur": 4.346,
    "args": {
      "External id": 2424,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 883,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 883, "pid": 537, "tid": 537, "ts": 7429564835653.662,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835666.644, "dur": 4.675,
    "args": {
      "External id": 2426,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 884,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 884, "pid": 537, "tid": 537, "ts": 7429564835666.644,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835709.193, "dur": 0.264,
    "args": {
      "External id": 2434,      "cid": 370, "correlation": 885
    }
  },
  {
    "ph": "f", "id": 885, "pid": 537, "tid": 537, "ts": 7429564835709.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835712.675, "dur": 0.163,
    "args": {
      "External id": 2434,      "cid": 370, "correlation": 886
    }
  },
  {
    "ph": "f", "id": 886, "pid": 537, "tid": 537, "ts": 7429564835712.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835715.076, "dur": 0.172,
    "args": {
      "External id": 2434,      "cid": 370, "correlation": 887
    }
  },
  {
    "ph": "f", "id": 887, "pid": 537, "tid": 537, "ts": 7429564835715.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835717.209, "dur": 6.278,
    "args": {
      "External id": 2434,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 888,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 888, "pid": 537, "tid": 537, "ts": 7429564835717.209,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835758.038, "dur": 0.190,
    "args": {
      "External id": 2447,      "cid": 370, "correlation": 889
    }
  },
  {
    "ph": "f", "id": 889, "pid": 537, "tid": 537, "ts": 7429564835758.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835761.203, "dur": 0.158,
    "args": {
      "External id": 2447,      "cid": 370, "correlation": 890
    }
  },
  {
    "ph": "f", "id": 890, "pid": 537, "tid": 537, "ts": 7429564835761.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835763.273, "dur": 0.138,
    "args": {
      "External id": 2447,      "cid": 370, "correlation": 891
    }
  },
  {
    "ph": "f", "id": 891, "pid": 537, "tid": 537, "ts": 7429564835763.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835767.384, "dur": 6.122,
    "args": {
      "External id": 2447,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 892,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 892, "pid": 537, "tid": 537, "ts": 7429564835767.384,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835773.868, "dur": 3.970,
    "args": {
      "External id": 2447,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 893,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 893, "pid": 537, "tid": 537, "ts": 7429564835773.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835811.409, "dur": 0.210,
    "args": {
      "External id": 2460,      "cid": 370, "correlation": 894
    }
  },
  {
    "ph": "f", "id": 894, "pid": 537, "tid": 537, "ts": 7429564835811.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835814.291, "dur": 0.149,
    "args": {
      "External id": 2460,      "cid": 370, "correlation": 895
    }
  },
  {
    "ph": "f", "id": 895, "pid": 537, "tid": 537, "ts": 7429564835814.291,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564835816.287, "dur": 0.138,
    "args": {
      "External id": 2460,      "cid": 370, "correlation": 896
    }
  },
  {
    "ph": "f", "id": 896, "pid": 537, "tid": 537, "ts": 7429564835816.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835819.818, "dur": 4.104,
    "args": {
      "External id": 2460,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 897,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 897, "pid": 537, "tid": 537, "ts": 7429564835819.818,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835824.391, "dur": 3.761,
    "args": {
      "External id": 2460,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 898,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 898, "pid": 537, "tid": 537, "ts": 7429564835824.391,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835853.750, "dur": 4.465,
    "args": {
      "External id": 2470,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 899,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 899, "pid": 537, "tid": 537, "ts": 7429564835853.750,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835879.974, "dur": 5.896,
    "args": {
      "External id": 2475,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 900,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 900, "pid": 537, "tid": 537, "ts": 7429564835879.974,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564835896.060, "dur": 0.268,
    "args": {
      "External id": 2476,      "cid": 290, "correlation": 901
    }
  },
  {
    "ph": "f", "id": 901, "pid": 537, "tid": 537, "ts": 7429564835896.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564835896.706, "dur": 37.480,
    "args": {
      "External id": 2476,            "bytes": 2097152, "cid": 108, "correlation": 902, "ptr": "0x5569ba9ecb70"
    }
  },
  {
    "ph": "s", "id": 902, "pid": 537, "tid": 537, "ts": 7429564835896.706,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835939.146, "dur": 7.391,
    "args": {
      "External id": 2476,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 903,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 903, "pid": 537, "tid": 537, "ts": 7429564835939.146,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835955.572, "dur": 3.998,
    "args": {
      "External id": 2477,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 904,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 904, "pid": 537, "tid": 537, "ts": 7429564835955.572,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835965.243, "dur": 4.167,
    "args": {
      "External id": 2478,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 905,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 905, "pid": 537, "tid": 537, "ts": 7429564835965.243,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835975.823, "dur": 4.083,
    "args": {
      "External id": 2479,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 906,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 906, "pid": 537, "tid": 537, "ts": 7429564835975.823,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564835994.169, "dur": 3.263,
    "args": {
      "External id": 2484,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 907,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 907, "pid": 537, "tid": 537, "ts": 7429564835994.169,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836005.978, "dur": 6.351,
    "args": {
      "External id": 2485,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 908,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 908, "pid": 537, "tid": 537, "ts": 7429564836005.978,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836017.999, "dur": 4.001,
    "args": {
      "External id": 2486,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 909,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 909, "pid": 537, "tid": 537, "ts": 7429564836017.999,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836026.379, "dur": 3.733,
    "args": {
      "External id": 2487,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 910,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 910, "pid": 537, "tid": 537, "ts": 7429564836026.379,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836089.214, "dur": 4.489,
    "args": {
      "External id": 2504,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 911,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 911, "pid": 537, "tid": 537, "ts": 7429564836089.214,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836125.153, "dur": 3.445,
    "args": {
      "External id": 2522,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 912,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 912, "pid": 537, "tid": 537, "ts": 7429564836125.153,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836164.853, "dur": 0.361,
    "args": {
      "External id": 2535,      "cid": 370, "correlation": 913
    }
  },
  {
    "ph": "f", "id": 913, "pid": 537, "tid": 537, "ts": 7429564836164.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836169.699, "dur": 0.156,
    "args": {
      "External id": 2535,      "cid": 370, "correlation": 914
    }
  },
  {
    "ph": "f", "id": 914, "pid": 537, "tid": 537, "ts": 7429564836169.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836172.292, "dur": 0.178,
    "args": {
      "External id": 2535,      "cid": 370, "correlation": 915
    }
  },
  {
    "ph": "f", "id": 915, "pid": 537, "tid": 537, "ts": 7429564836172.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836175.414, "dur": 5.869,
    "args": {
      "External id": 2535,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 916,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 916, "pid": 537, "tid": 537, "ts": 7429564836175.414,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836193.134, "dur": 5.456,
    "args": {
      "External id": 2538,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 917,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 917, "pid": 537, "tid": 537, "ts": 7429564836193.134,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836213.142, "dur": 4.171,
    "args": {
      "External id": 2545,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 918,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 918, "pid": 537, "tid": 537, "ts": 7429564836213.142,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836229.392, "dur": 5.935,
    "args": {
      "External id": 2548,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 919,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 919, "pid": 537, "tid": 537, "ts": 7429564836229.392,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836309.951, "dur": 0.218,
    "args": {
      "External id": 2560,      "cid": 370, "correlation": 920
    }
  },
  {
    "ph": "f", "id": 920, "pid": 537, "tid": 537, "ts": 7429564836309.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836313.087, "dur": 0.143,
    "args": {
      "External id": 2560,      "cid": 370, "correlation": 921
    }
  },
  {
    "ph": "f", "id": 921, "pid": 537, "tid": 537, "ts": 7429564836313.087,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836315.945, "dur": 0.168,
    "args": {
      "External id": 2560,      "cid": 370, "correlation": 922
    }
  },
  {
    "ph": "f", "id": 922, "pid": 537, "tid": 537, "ts": 7429564836315.945,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836318.222, "dur": 5.287,
    "args": {
      "External id": 2560,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 923,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 923, "pid": 537, "tid": 537, "ts": 7429564836318.222,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836340.304, "dur": 3.796,
    "args": {
      "External id": 2569,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 924,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 924, "pid": 537, "tid": 537, "ts": 7429564836340.304,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836388.672, "dur": 0.226,
    "args": {
      "External id": 2579,      "cid": 370, "correlation": 925
    }
  },
  {
    "ph": "f", "id": 925, "pid": 537, "tid": 537, "ts": 7429564836388.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836392.029, "dur": 0.183,
    "args": {
      "External id": 2579,      "cid": 370, "correlation": 926
    }
  },
  {
    "ph": "f", "id": 926, "pid": 537, "tid": 537, "ts": 7429564836392.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836394.187, "dur": 0.311,
    "args": {
      "External id": 2579,      "cid": 370, "correlation": 927
    }
  },
  {
    "ph": "f", "id": 927, "pid": 537, "tid": 537, "ts": 7429564836394.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836396.147, "dur": 5.039,
    "args": {
      "External id": 2579,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 928,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 928, "pid": 537, "tid": 537, "ts": 7429564836396.147,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836413.595, "dur": 5.541,
    "args": {
      "External id": 2582,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 929,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 929, "pid": 537, "tid": 537, "ts": 7429564836413.595,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836438.337, "dur": 5.446,
    "args": {
      "External id": 2584,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 930,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 930, "pid": 537, "tid": 537, "ts": 7429564836438.337,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836454.141, "dur": 4.760,
    "args": {
      "External id": 2587,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 931,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 931, "pid": 537, "tid": 537, "ts": 7429564836454.141,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836467.142, "dur": 4.437,
    "args": {
      "External id": 2588,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 932,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 932, "pid": 537, "tid": 537, "ts": 7429564836467.142,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836479.427, "dur": 4.558,
    "args": {
      "External id": 2589,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 933,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 933, "pid": 537, "tid": 537, "ts": 7429564836479.427,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836490.718, "dur": 5.050,
    "args": {
      "External id": 2590,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 934,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 934, "pid": 537, "tid": 537, "ts": 7429564836490.718,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836504.547, "dur": 5.127,
    "args": {
      "External id": 2592,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 935,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 935, "pid": 537, "tid": 537, "ts": 7429564836504.547,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836543.114, "dur": 0.217,
    "args": {
      "External id": 2600,      "cid": 370, "correlation": 936
    }
  },
  {
    "ph": "f", "id": 936, "pid": 537, "tid": 537, "ts": 7429564836543.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836546.791, "dur": 0.159,
    "args": {
      "External id": 2600,      "cid": 370, "correlation": 937
    }
  },
  {
    "ph": "f", "id": 937, "pid": 537, "tid": 537, "ts": 7429564836546.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836549.449, "dur": 0.171,
    "args": {
      "External id": 2600,      "cid": 370, "correlation": 938
    }
  },
  {
    "ph": "f", "id": 938, "pid": 537, "tid": 537, "ts": 7429564836549.449,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836557.157, "dur": 5.504,
    "args": {
      "External id": 2600,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 939,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 939, "pid": 537, "tid": 537, "ts": 7429564836557.157,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836563.196, "dur": 3.759,
    "args": {
      "External id": 2600,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 940,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 940, "pid": 537, "tid": 537, "ts": 7429564836563.196,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564836583.018, "dur": 0.322,
    "args": {
      "External id": 2603,      "cid": 290, "correlation": 941
    }
  },
  {
    "ph": "f", "id": 941, "pid": 537, "tid": 537, "ts": 7429564836583.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564836583.803, "dur": 51.496,
    "args": {
      "External id": 2603,            "bytes": 2097152, "cid": 108, "correlation": 942, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 942, "pid": 537, "tid": 537, "ts": 7429564836583.803,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836638.537, "dur": 5.018,
    "args": {
      "External id": 2603,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 943,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 943, "pid": 537, "tid": 537, "ts": 7429564836638.537,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836675.983, "dur": 0.214,
    "args": {
      "External id": 2611,      "cid": 370, "correlation": 944
    }
  },
  {
    "ph": "f", "id": 944, "pid": 537, "tid": 537, "ts": 7429564836675.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836679.158, "dur": 0.140,
    "args": {
      "External id": 2611,      "cid": 370, "correlation": 945
    }
  },
  {
    "ph": "f", "id": 945, "pid": 537, "tid": 537, "ts": 7429564836679.158,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836681.420, "dur": 0.147,
    "args": {
      "External id": 2611,      "cid": 370, "correlation": 946
    }
  },
  {
    "ph": "f", "id": 946, "pid": 537, "tid": 537, "ts": 7429564836681.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836686.690, "dur": 4.917,
    "args": {
      "External id": 2611,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 947,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 947, "pid": 537, "tid": 537, "ts": 7429564836686.690,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836692.132, "dur": 3.748,
    "args": {
      "External id": 2611,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 948,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 948, "pid": 537, "tid": 537, "ts": 7429564836692.132,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564836704.611, "dur": 0.186,
    "args": {
      "External id": 2614,      "cid": 290, "correlation": 949
    }
  },
  {
    "ph": "f", "id": 949, "pid": 537, "tid": 537, "ts": 7429564836704.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564836705.122, "dur": 29.087,
    "args": {
      "External id": 2614,            "bytes": 2097152, "cid": 108, "correlation": 950, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 950, "pid": 537, "tid": 537, "ts": 7429564836705.122,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836737.409, "dur": 4.945,
    "args": {
      "External id": 2614,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 951,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 951, "pid": 537, "tid": 537, "ts": 7429564836737.409,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836772.483, "dur": 0.182,
    "args": {
      "External id": 2622,      "cid": 370, "correlation": 952
    }
  },
  {
    "ph": "f", "id": 952, "pid": 537, "tid": 537, "ts": 7429564836772.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836775.819, "dur": 0.145,
    "args": {
      "External id": 2622,      "cid": 370, "correlation": 953
    }
  },
  {
    "ph": "f", "id": 953, "pid": 537, "tid": 537, "ts": 7429564836775.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836778.145, "dur": 0.142,
    "args": {
      "External id": 2622,      "cid": 370, "correlation": 954
    }
  },
  {
    "ph": "f", "id": 954, "pid": 537, "tid": 537, "ts": 7429564836778.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836782.510, "dur": 5.178,
    "args": {
      "External id": 2622,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 955,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 955, "pid": 537, "tid": 537, "ts": 7429564836782.510,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836788.168, "dur": 3.191,
    "args": {
      "External id": 2622,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 956,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 956, "pid": 537, "tid": 537, "ts": 7429564836788.168,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836801.405, "dur": 4.639,
    "args": {
      "External id": 2625,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 957,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 957, "pid": 537, "tid": 537, "ts": 7429564836801.405,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836836.203, "dur": 5.098,
    "args": {
      "External id": 2627,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 958,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 958, "pid": 537, "tid": 537, "ts": 7429564836836.203,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836850.746, "dur": 5.211,
    "args": {
      "External id": 2630,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 959,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 959, "pid": 537, "tid": 537, "ts": 7429564836850.746,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836864.951, "dur": 4.643,
    "args": {
      "External id": 2631,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 960,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 960, "pid": 537, "tid": 537, "ts": 7429564836864.951,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836874.881, "dur": 4.519,
    "args": {
      "External id": 2632,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 961,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 961, "pid": 537, "tid": 537, "ts": 7429564836874.881,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836887.066, "dur": 3.987,
    "args": {
      "External id": 2633,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 962,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 962, "pid": 537, "tid": 537, "ts": 7429564836887.066,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836898.619, "dur": 5.014,
    "args": {
      "External id": 2635,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 963,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 963, "pid": 537, "tid": 537, "ts": 7429564836898.619,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836943.583, "dur": 0.269,
    "args": {
      "External id": 2643,      "cid": 370, "correlation": 964
    }
  },
  {
    "ph": "f", "id": 964, "pid": 537, "tid": 537, "ts": 7429564836943.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836946.787, "dur": 0.157,
    "args": {
      "External id": 2643,      "cid": 370, "correlation": 965
    }
  },
  {
    "ph": "f", "id": 965, "pid": 537, "tid": 537, "ts": 7429564836946.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836949.034, "dur": 0.162,
    "args": {
      "External id": 2643,      "cid": 370, "correlation": 966
    }
  },
  {
    "ph": "f", "id": 966, "pid": 537, "tid": 537, "ts": 7429564836949.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564836951.084, "dur": 4.387,
    "args": {
      "External id": 2643,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 967,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 967, "pid": 537, "tid": 537, "ts": 7429564836951.084,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836993.686, "dur": 0.220,
    "args": {
      "External id": 2656,      "cid": 370, "correlation": 968
    }
  },
  {
    "ph": "f", "id": 968, "pid": 537, "tid": 537, "ts": 7429564836993.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836996.554, "dur": 0.155,
    "args": {
      "External id": 2656,      "cid": 370, "correlation": 969
    }
  },
  {
    "ph": "f", "id": 969, "pid": 537, "tid": 537, "ts": 7429564836996.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564836998.545, "dur": 0.181,
    "args": {
      "External id": 2656,      "cid": 370, "correlation": 970
    }
  },
  {
    "ph": "f", "id": 970, "pid": 537, "tid": 537, "ts": 7429564836998.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837002.583, "dur": 5.943,
    "args": {
      "External id": 2656,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 971,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 971, "pid": 537, "tid": 537, "ts": 7429564837002.583,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837009.136, "dur": 5.245,
    "args": {
      "External id": 2656,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 972,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 972, "pid": 537, "tid": 537, "ts": 7429564837009.136,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837051.039, "dur": 0.174,
    "args": {
      "External id": 2669,      "cid": 370, "correlation": 973
    }
  },
  {
    "ph": "f", "id": 973, "pid": 537, "tid": 537, "ts": 7429564837051.039,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837053.909, "dur": 0.141,
    "args": {
      "External id": 2669,      "cid": 370, "correlation": 974
    }
  },
  {
    "ph": "f", "id": 974, "pid": 537, "tid": 537, "ts": 7429564837053.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837055.884, "dur": 0.140,
    "args": {
      "External id": 2669,      "cid": 370, "correlation": 975
    }
  },
  {
    "ph": "f", "id": 975, "pid": 537, "tid": 537, "ts": 7429564837055.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837059.301, "dur": 4.830,
    "args": {
      "External id": 2669,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 976,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 976, "pid": 537, "tid": 537, "ts": 7429564837059.301,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837064.469, "dur": 3.166,
    "args": {
      "External id": 2669,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 977,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 977, "pid": 537, "tid": 537, "ts": 7429564837064.469,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837091.251, "dur": 4.275,
    "args": {
      "External id": 2679,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 978,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 978, "pid": 537, "tid": 537, "ts": 7429564837091.251,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837118.400, "dur": 5.140,
    "args": {
      "External id": 2684,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 979,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 979, "pid": 537, "tid": 537, "ts": 7429564837118.400,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837138.560, "dur": 6.768,
    "args": {
      "External id": 2685,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 980,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 980, "pid": 537, "tid": 537, "ts": 7429564837138.560,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837152.463, "dur": 5.566,
    "args": {
      "External id": 2686,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 981,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 981, "pid": 537, "tid": 537, "ts": 7429564837152.463,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837164.532, "dur": 4.938,
    "args": {
      "External id": 2687,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 982,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 982, "pid": 537, "tid": 537, "ts": 7429564837164.532,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837175.044, "dur": 3.414,
    "args": {
      "External id": 2688,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 983,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 983, "pid": 537, "tid": 537, "ts": 7429564837175.044,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837190.824, "dur": 3.982,
    "args": {
      "External id": 2693,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 984,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 984, "pid": 537, "tid": 537, "ts": 7429564837190.824,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837201.045, "dur": 4.568,
    "args": {
      "External id": 2694,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 985,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 985, "pid": 537, "tid": 537, "ts": 7429564837201.045,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837211.795, "dur": 4.025,
    "args": {
      "External id": 2695,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 986,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 986, "pid": 537, "tid": 537, "ts": 7429564837211.795,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837220.601, "dur": 4.357,
    "args": {
      "External id": 2696,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 987,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 987, "pid": 537, "tid": 537, "ts": 7429564837220.601,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837284.073, "dur": 5.229,
    "args": {
      "External id": 2713,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 988,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 988, "pid": 537, "tid": 537, "ts": 7429564837284.073,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837318.861, "dur": 4.305,
    "args": {
      "External id": 2731,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 989,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 989, "pid": 537, "tid": 537, "ts": 7429564837318.861,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837356.314, "dur": 0.311,
    "args": {
      "External id": 2744,      "cid": 370, "correlation": 990
    }
  },
  {
    "ph": "f", "id": 990, "pid": 537, "tid": 537, "ts": 7429564837356.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837360.524, "dur": 0.159,
    "args": {
      "External id": 2744,      "cid": 370, "correlation": 991
    }
  },
  {
    "ph": "f", "id": 991, "pid": 537, "tid": 537, "ts": 7429564837360.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837362.860, "dur": 0.163,
    "args": {
      "External id": 2744,      "cid": 370, "correlation": 992
    }
  },
  {
    "ph": "f", "id": 992, "pid": 537, "tid": 537, "ts": 7429564837362.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837365.204, "dur": 4.198,
    "args": {
      "External id": 2744,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 993,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 993, "pid": 537, "tid": 537, "ts": 7429564837365.204,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837380.467, "dur": 4.312,
    "args": {
      "External id": 2747,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 994,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 994, "pid": 537, "tid": 537, "ts": 7429564837380.467,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837399.322, "dur": 4.502,
    "args": {
      "External id": 2754,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 995,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 995, "pid": 537, "tid": 537, "ts": 7429564837399.322,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837414.672, "dur": 6.111,
    "args": {
      "External id": 2757,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 996,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 996, "pid": 537, "tid": 537, "ts": 7429564837414.672,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837463.639, "dur": 0.197,
    "args": {
      "External id": 2769,      "cid": 370, "correlation": 997
    }
  },
  {
    "ph": "f", "id": 997, "pid": 537, "tid": 537, "ts": 7429564837463.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837466.856, "dur": 0.157,
    "args": {
      "External id": 2769,      "cid": 370, "correlation": 998
    }
  },
  {
    "ph": "f", "id": 998, "pid": 537, "tid": 537, "ts": 7429564837466.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837469.816, "dur": 0.144,
    "args": {
      "External id": 2769,      "cid": 370, "correlation": 999
    }
  },
  {
    "ph": "f", "id": 999, "pid": 537, "tid": 537, "ts": 7429564837469.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837471.752, "dur": 5.985,
    "args": {
      "External id": 2769,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1000,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1000, "pid": 537, "tid": 537, "ts": 7429564837471.752,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837493.718, "dur": 3.745,
    "args": {
      "External id": 2778,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1001,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1001, "pid": 537, "tid": 537, "ts": 7429564837493.718,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837538.694, "dur": 0.189,
    "args": {
      "External id": 2788,      "cid": 370, "correlation": 1002
    }
  },
  {
    "ph": "f", "id": 1002, "pid": 537, "tid": 537, "ts": 7429564837538.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837541.986, "dur": 0.154,
    "args": {
      "External id": 2788,      "cid": 370, "correlation": 1003
    }
  },
  {
    "ph": "f", "id": 1003, "pid": 537, "tid": 537, "ts": 7429564837541.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837544.415, "dur": 0.180,
    "args": {
      "External id": 2788,      "cid": 370, "correlation": 1004
    }
  },
  {
    "ph": "f", "id": 1004, "pid": 537, "tid": 537, "ts": 7429564837544.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837546.183, "dur": 4.448,
    "args": {
      "External id": 2788,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1005,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1005, "pid": 537, "tid": 537, "ts": 7429564837546.183,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837562.270, "dur": 4.903,
    "args": {
      "External id": 2791,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1006,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1006, "pid": 537, "tid": 537, "ts": 7429564837562.270,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837583.503, "dur": 5.426,
    "args": {
      "External id": 2793,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1007,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1007, "pid": 537, "tid": 537, "ts": 7429564837583.503,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837598.234, "dur": 6.015,
    "args": {
      "External id": 2796,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1008,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1008, "pid": 537, "tid": 537, "ts": 7429564837598.234,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837612.555, "dur": 4.170,
    "args": {
      "External id": 2797,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1009,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1009, "pid": 537, "tid": 537, "ts": 7429564837612.555,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837622.407, "dur": 3.779,
    "args": {
      "External id": 2798,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1010,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1010, "pid": 537, "tid": 537, "ts": 7429564837622.407,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837632.622, "dur": 4.053,
    "args": {
      "External id": 2799,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1011,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1011, "pid": 537, "tid": 537, "ts": 7429564837632.622,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564837645.030, "dur": 0.274,
    "args": {
      "External id": 2801,      "cid": 290, "correlation": 1012
    }
  },
  {
    "ph": "f", "id": 1012, "pid": 537, "tid": 537, "ts": 7429564837645.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564837645.711, "dur": 50.393,
    "args": {
      "External id": 2801,            "bytes": 2097152, "cid": 108, "correlation": 1013, "ptr": "0x100000002"
    }
  },
  {
    "ph": "s", "id": 1013, "pid": 537, "tid": 537, "ts": 7429564837645.711,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837698.756, "dur": 4.421,
    "args": {
      "External id": 2801,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1014,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1014, "pid": 537, "tid": 537, "ts": 7429564837698.756,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837742.741, "dur": 0.227,
    "args": {
      "External id": 2809,      "cid": 370, "correlation": 1015
    }
  },
  {
    "ph": "f", "id": 1015, "pid": 537, "tid": 537, "ts": 7429564837742.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837746.530, "dur": 0.146,
    "args": {
      "External id": 2809,      "cid": 370, "correlation": 1016
    }
  },
  {
    "ph": "f", "id": 1016, "pid": 537, "tid": 537, "ts": 7429564837746.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837748.986, "dur": 0.139,
    "args": {
      "External id": 2809,      "cid": 370, "correlation": 1017
    }
  },
  {
    "ph": "f", "id": 1017, "pid": 537, "tid": 537, "ts": 7429564837748.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837756.468, "dur": 5.383,
    "args": {
      "External id": 2809,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1018,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1018, "pid": 537, "tid": 537, "ts": 7429564837756.468,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837762.200, "dur": 4.216,
    "args": {
      "External id": 2809,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1019,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1019, "pid": 537, "tid": 537, "ts": 7429564837762.200,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837784.106, "dur": 4.510,
    "args": {
      "External id": 2812,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1020,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1020, "pid": 537, "tid": 537, "ts": 7429564837784.106,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564837805.936, "dur": 0.309,
    "args": {
      "External id": 2820,      "cid": 290, "correlation": 1021
    }
  },
  {
    "ph": "f", "id": 1021, "pid": 537, "tid": 537, "ts": 7429564837805.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564837806.678, "dur": 33.209,
    "args": {
      "External id": 2820,            "bytes": 2097152, "cid": 108, "correlation": 1022, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1022, "pid": 537, "tid": 537, "ts": 7429564837806.678,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837851.344, "dur": 0.217,
    "args": {
      "External id": 2820,      "cid": 370, "correlation": 1023
    }
  },
  {
    "ph": "f", "id": 1023, "pid": 537, "tid": 537, "ts": 7429564837851.344,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837854.513, "dur": 0.151,
    "args": {
      "External id": 2820,      "cid": 370, "correlation": 1024
    }
  },
  {
    "ph": "f", "id": 1024, "pid": 537, "tid": 537, "ts": 7429564837854.513,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837857.020, "dur": 0.141,
    "args": {
      "External id": 2820,      "cid": 370, "correlation": 1025
    }
  },
  {
    "ph": "f", "id": 1025, "pid": 537, "tid": 537, "ts": 7429564837857.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837868.030, "dur": 5.757,
    "args": {
      "External id": 2820,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1026,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1026, "pid": 537, "tid": 537, "ts": 7429564837868.030,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837874.093, "dur": 3.606,
    "args": {
      "External id": 2820,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1027,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1027, "pid": 537, "tid": 537, "ts": 7429564837874.093,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837887.683, "dur": 4.243,
    "args": {
      "External id": 2823,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1028,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1028, "pid": 537, "tid": 537, "ts": 7429564837887.683,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837914.903, "dur": 0.183,
    "args": {
      "External id": 2831,      "cid": 370, "correlation": 1029
    }
  },
  {
    "ph": "f", "id": 1029, "pid": 537, "tid": 537, "ts": 7429564837914.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837917.550, "dur": 0.144,
    "args": {
      "External id": 2831,      "cid": 370, "correlation": 1030
    }
  },
  {
    "ph": "f", "id": 1030, "pid": 537, "tid": 537, "ts": 7429564837917.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564837919.418, "dur": 0.157,
    "args": {
      "External id": 2831,      "cid": 370, "correlation": 1031
    }
  },
  {
    "ph": "f", "id": 1031, "pid": 537, "tid": 537, "ts": 7429564837919.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837923.223, "dur": 4.908,
    "args": {
      "External id": 2831,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1032,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1032, "pid": 537, "tid": 537, "ts": 7429564837923.223,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837928.447, "dur": 4.041,
    "args": {
      "External id": 2831,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1033,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1033, "pid": 537, "tid": 537, "ts": 7429564837928.447,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837942.256, "dur": 5.006,
    "args": {
      "External id": 2834,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1034,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1034, "pid": 537, "tid": 537, "ts": 7429564837942.256,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837975.289, "dur": 4.208,
    "args": {
      "External id": 2836,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1035,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1035, "pid": 537, "tid": 537, "ts": 7429564837975.289,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564837988.601, "dur": 5.271,
    "args": {
      "External id": 2839,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1036,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1036, "pid": 537, "tid": 537, "ts": 7429564837988.601,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838001.735, "dur": 3.664,
    "args": {
      "External id": 2840,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1037,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1037, "pid": 537, "tid": 537, "ts": 7429564838001.735,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838010.369, "dur": 4.217,
    "args": {
      "External id": 2841,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1038,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1038, "pid": 537, "tid": 537, "ts": 7429564838010.369,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838020.714, "dur": 4.447,
    "args": {
      "External id": 2842,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1039,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1039, "pid": 537, "tid": 537, "ts": 7429564838020.714,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838032.159, "dur": 5.031,
    "args": {
      "External id": 2844,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1040,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1040, "pid": 537, "tid": 537, "ts": 7429564838032.159,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564838066.702, "dur": 0.246,
    "args": {
      "External id": 2852,      "cid": 290, "correlation": 1041
    }
  },
  {
    "ph": "f", "id": 1041, "pid": 537, "tid": 537, "ts": 7429564838066.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564838067.286, "dur": 53.062,
    "args": {
      "External id": 2852,            "bytes": 2097152, "cid": 108, "correlation": 1042, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1042, "pid": 537, "tid": 537, "ts": 7429564838067.286,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838138.347, "dur": 0.374,
    "args": {
      "External id": 2852,      "cid": 370, "correlation": 1043
    }
  },
  {
    "ph": "f", "id": 1043, "pid": 537, "tid": 537, "ts": 7429564838138.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838143.602, "dur": 0.154,
    "args": {
      "External id": 2852,      "cid": 370, "correlation": 1044
    }
  },
  {
    "ph": "f", "id": 1044, "pid": 537, "tid": 537, "ts": 7429564838143.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838146.566, "dur": 0.152,
    "args": {
      "External id": 2852,      "cid": 370, "correlation": 1045
    }
  },
  {
    "ph": "f", "id": 1045, "pid": 537, "tid": 537, "ts": 7429564838146.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838149.933, "dur": 5.609,
    "args": {
      "External id": 2852,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1046,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1046, "pid": 537, "tid": 537, "ts": 7429564838149.933,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838204.531, "dur": 0.208,
    "args": {
      "External id": 2865,      "cid": 370, "correlation": 1047
    }
  },
  {
    "ph": "f", "id": 1047, "pid": 537, "tid": 537, "ts": 7429564838204.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838207.613, "dur": 0.166,
    "args": {
      "External id": 2865,      "cid": 370, "correlation": 1048
    }
  },
  {
    "ph": "f", "id": 1048, "pid": 537, "tid": 537, "ts": 7429564838207.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838209.806, "dur": 0.146,
    "args": {
      "External id": 2865,      "cid": 370, "correlation": 1049
    }
  },
  {
    "ph": "f", "id": 1049, "pid": 537, "tid": 537, "ts": 7429564838209.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838215.355, "dur": 5.525,
    "args": {
      "External id": 2865,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1050,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1050, "pid": 537, "tid": 537, "ts": 7429564838215.355,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838221.251, "dur": 4.224,
    "args": {
      "External id": 2865,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1051,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1051, "pid": 537, "tid": 537, "ts": 7429564838221.251,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838263.619, "dur": 0.199,
    "args": {
      "External id": 2878,      "cid": 370, "correlation": 1052
    }
  },
  {
    "ph": "f", "id": 1052, "pid": 537, "tid": 537, "ts": 7429564838263.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838266.193, "dur": 0.138,
    "args": {
      "External id": 2878,      "cid": 370, "correlation": 1053
    }
  },
  {
    "ph": "f", "id": 1053, "pid": 537, "tid": 537, "ts": 7429564838266.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838268.109, "dur": 0.141,
    "args": {
      "External id": 2878,      "cid": 370, "correlation": 1054
    }
  },
  {
    "ph": "f", "id": 1054, "pid": 537, "tid": 537, "ts": 7429564838268.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838271.643, "dur": 4.707,
    "args": {
      "External id": 2878,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1055,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1055, "pid": 537, "tid": 537, "ts": 7429564838271.643,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838276.627, "dur": 4.540,
    "args": {
      "External id": 2878,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1056,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1056, "pid": 537, "tid": 537, "ts": 7429564838276.627,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838307.202, "dur": 5.646,
    "args": {
      "External id": 2888,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1057,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1057, "pid": 537, "tid": 537, "ts": 7429564838307.202,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838337.484, "dur": 6.101,
    "args": {
      "External id": 2893,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1058,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1058, "pid": 537, "tid": 537, "ts": 7429564838337.484,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838358.931, "dur": 8.169,
    "args": {
      "External id": 2894,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1059,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1059, "pid": 537, "tid": 537, "ts": 7429564838358.931,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838374.853, "dur": 3.800,
    "args": {
      "External id": 2895,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1060,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1060, "pid": 537, "tid": 537, "ts": 7429564838374.853,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838385.569, "dur": 5.360,
    "args": {
      "External id": 2896,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1061,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1061, "pid": 537, "tid": 537, "ts": 7429564838385.569,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838396.682, "dur": 4.278,
    "args": {
      "External id": 2897,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1062,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1062, "pid": 537, "tid": 537, "ts": 7429564838396.682,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838415.005, "dur": 4.466,
    "args": {
      "External id": 2902,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1063,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1063, "pid": 537, "tid": 537, "ts": 7429564838415.005,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838425.499, "dur": 5.402,
    "args": {
      "External id": 2903,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1064,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1064, "pid": 537, "tid": 537, "ts": 7429564838425.499,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838437.038, "dur": 4.320,
    "args": {
      "External id": 2904,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1065,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1065, "pid": 537, "tid": 537, "ts": 7429564838437.038,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838446.486, "dur": 3.502,
    "args": {
      "External id": 2905,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1066,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1066, "pid": 537, "tid": 537, "ts": 7429564838446.486,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838512.924, "dur": 4.870,
    "args": {
      "External id": 2922,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1067,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1067, "pid": 537, "tid": 537, "ts": 7429564838512.924,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838548.060, "dur": 4.538,
    "args": {
      "External id": 2940,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1068,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1068, "pid": 537, "tid": 537, "ts": 7429564838548.060,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838587.851, "dur": 0.253,
    "args": {
      "External id": 2953,      "cid": 370, "correlation": 1069
    }
  },
  {
    "ph": "f", "id": 1069, "pid": 537, "tid": 537, "ts": 7429564838587.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838592.070, "dur": 0.156,
    "args": {
      "External id": 2953,      "cid": 370, "correlation": 1070
    }
  },
  {
    "ph": "f", "id": 1070, "pid": 537, "tid": 537, "ts": 7429564838592.070,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838594.584, "dur": 0.181,
    "args": {
      "External id": 2953,      "cid": 370, "correlation": 1071
    }
  },
  {
    "ph": "f", "id": 1071, "pid": 537, "tid": 537, "ts": 7429564838594.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838597.051, "dur": 6.039,
    "args": {
      "External id": 2953,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1072,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1072, "pid": 537, "tid": 537, "ts": 7429564838597.051,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838615.590, "dur": 3.995,
    "args": {
      "External id": 2956,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1073,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1073, "pid": 537, "tid": 537, "ts": 7429564838615.590,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838633.633, "dur": 3.907,
    "args": {
      "External id": 2963,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1074,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1074, "pid": 537, "tid": 537, "ts": 7429564838633.633,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838649.829, "dur": 5.552,
    "args": {
      "External id": 2966,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1075,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1075, "pid": 537, "tid": 537, "ts": 7429564838649.829,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838698.877, "dur": 0.190,
    "args": {
      "External id": 2978,      "cid": 370, "correlation": 1076
    }
  },
  {
    "ph": "f", "id": 1076, "pid": 537, "tid": 537, "ts": 7429564838698.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838701.849, "dur": 0.146,
    "args": {
      "External id": 2978,      "cid": 370, "correlation": 1077
    }
  },
  {
    "ph": "f", "id": 1077, "pid": 537, "tid": 537, "ts": 7429564838701.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838704.711, "dur": 0.170,
    "args": {
      "External id": 2978,      "cid": 370, "correlation": 1078
    }
  },
  {
    "ph": "f", "id": 1078, "pid": 537, "tid": 537, "ts": 7429564838704.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838706.828, "dur": 4.938,
    "args": {
      "External id": 2978,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1079,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1079, "pid": 537, "tid": 537, "ts": 7429564838706.828,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838728.421, "dur": 4.554,
    "args": {
      "External id": 2987,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1080,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1080, "pid": 537, "tid": 537, "ts": 7429564838728.421,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838774.554, "dur": 0.186,
    "args": {
      "External id": 2997,      "cid": 370, "correlation": 1081
    }
  },
  {
    "ph": "f", "id": 1081, "pid": 537, "tid": 537, "ts": 7429564838774.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838777.505, "dur": 0.154,
    "args": {
      "External id": 2997,      "cid": 370, "correlation": 1082
    }
  },
  {
    "ph": "f", "id": 1082, "pid": 537, "tid": 537, "ts": 7429564838777.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838779.705, "dur": 0.175,
    "args": {
      "External id": 2997,      "cid": 370, "correlation": 1083
    }
  },
  {
    "ph": "f", "id": 1083, "pid": 537, "tid": 537, "ts": 7429564838779.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838781.486, "dur": 4.744,
    "args": {
      "External id": 2997,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1084,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1084, "pid": 537, "tid": 537, "ts": 7429564838781.486,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838798.315, "dur": 5.119,
    "args": {
      "External id": 3000,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1085,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1085, "pid": 537, "tid": 537, "ts": 7429564838798.315,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838821.669, "dur": 4.937,
    "args": {
      "External id": 3002,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1086,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1086, "pid": 537, "tid": 537, "ts": 7429564838821.669,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838836.478, "dur": 7.274,
    "args": {
      "External id": 3005,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1087,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1087, "pid": 537, "tid": 537, "ts": 7429564838836.478,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838851.768, "dur": 4.067,
    "args": {
      "External id": 3006,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1088,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1088, "pid": 537, "tid": 537, "ts": 7429564838851.768,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838862.633, "dur": 5.044,
    "args": {
      "External id": 3007,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1089,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1089, "pid": 537, "tid": 537, "ts": 7429564838862.633,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838875.212, "dur": 4.553,
    "args": {
      "External id": 3008,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1090,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1090, "pid": 537, "tid": 537, "ts": 7429564838875.212,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838887.389, "dur": 5.283,
    "args": {
      "External id": 3010,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1091,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1091, "pid": 537, "tid": 537, "ts": 7429564838887.389,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564838918.979, "dur": 0.403,
    "args": {
      "External id": 3018,      "cid": 290, "correlation": 1092
    }
  },
  {
    "ph": "f", "id": 1092, "pid": 537, "tid": 537, "ts": 7429564838918.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564838919.889, "dur": 51.829,
    "args": {
      "External id": 3018,            "bytes": 2097152, "cid": 108, "correlation": 1093, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1093, "pid": 537, "tid": 537, "ts": 7429564838919.889,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838983.800, "dur": 0.237,
    "args": {
      "External id": 3018,      "cid": 370, "correlation": 1094
    }
  },
  {
    "ph": "f", "id": 1094, "pid": 537, "tid": 537, "ts": 7429564838983.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838987.609, "dur": 0.140,
    "args": {
      "External id": 3018,      "cid": 370, "correlation": 1095
    }
  },
  {
    "ph": "f", "id": 1095, "pid": 537, "tid": 537, "ts": 7429564838987.609,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564838989.928, "dur": 0.160,
    "args": {
      "External id": 3018,      "cid": 370, "correlation": 1096
    }
  },
  {
    "ph": "f", "id": 1096, "pid": 537, "tid": 537, "ts": 7429564838989.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564838997.983, "dur": 5.646,
    "args": {
      "External id": 3018,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1097,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1097, "pid": 537, "tid": 537, "ts": 7429564838997.983,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839003.956, "dur": 3.692,
    "args": {
      "External id": 3018,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1098,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1098, "pid": 537, "tid": 537, "ts": 7429564839003.956,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839026.676, "dur": 4.314,
    "args": {
      "External id": 3021,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1099,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1099, "pid": 537, "tid": 537, "ts": 7429564839026.676,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564839050.393, "dur": 0.225,
    "args": {
      "External id": 3029,      "cid": 290, "correlation": 1100
    }
  },
  {
    "ph": "f", "id": 1100, "pid": 537, "tid": 537, "ts": 7429564839050.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564839050.945, "dur": 27.648,
    "args": {
      "External id": 3029,            "bytes": 2097152, "cid": 108, "correlation": 1101, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1101, "pid": 537, "tid": 537, "ts": 7429564839050.945,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839091.009, "dur": 0.197,
    "args": {
      "External id": 3029,      "cid": 370, "correlation": 1102
    }
  },
  {
    "ph": "f", "id": 1102, "pid": 537, "tid": 537, "ts": 7429564839091.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839094.264, "dur": 0.148,
    "args": {
      "External id": 3029,      "cid": 370, "correlation": 1103
    }
  },
  {
    "ph": "f", "id": 1103, "pid": 537, "tid": 537, "ts": 7429564839094.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839096.436, "dur": 0.142,
    "args": {
      "External id": 3029,      "cid": 370, "correlation": 1104
    }
  },
  {
    "ph": "f", "id": 1104, "pid": 537, "tid": 537, "ts": 7429564839096.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839100.900, "dur": 5.389,
    "args": {
      "External id": 3029,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1105,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1105, "pid": 537, "tid": 537, "ts": 7429564839100.900,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839106.611, "dur": 4.432,
    "args": {
      "External id": 3029,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1106,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1106, "pid": 537, "tid": 537, "ts": 7429564839106.611,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839121.123, "dur": 4.256,
    "args": {
      "External id": 3032,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1107,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1107, "pid": 537, "tid": 537, "ts": 7429564839121.123,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839150.886, "dur": 0.198,
    "args": {
      "External id": 3040,      "cid": 370, "correlation": 1108
    }
  },
  {
    "ph": "f", "id": 1108, "pid": 537, "tid": 537, "ts": 7429564839150.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839154.198, "dur": 0.150,
    "args": {
      "External id": 3040,      "cid": 370, "correlation": 1109
    }
  },
  {
    "ph": "f", "id": 1109, "pid": 537, "tid": 537, "ts": 7429564839154.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839156.113, "dur": 0.143,
    "args": {
      "External id": 3040,      "cid": 370, "correlation": 1110
    }
  },
  {
    "ph": "f", "id": 1110, "pid": 537, "tid": 537, "ts": 7429564839156.113,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839159.826, "dur": 4.535,
    "args": {
      "External id": 3040,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1111,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1111, "pid": 537, "tid": 537, "ts": 7429564839159.826,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839164.643, "dur": 3.197,
    "args": {
      "External id": 3040,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1112,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1112, "pid": 537, "tid": 537, "ts": 7429564839164.643,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839178.508, "dur": 5.249,
    "args": {
      "External id": 3043,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1113,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1113, "pid": 537, "tid": 537, "ts": 7429564839178.508,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839209.904, "dur": 4.529,
    "args": {
      "External id": 3045,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1114,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1114, "pid": 537, "tid": 537, "ts": 7429564839209.904,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839223.511, "dur": 5.511,
    "args": {
      "External id": 3048,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1115,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1115, "pid": 537, "tid": 537, "ts": 7429564839223.511,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839237.665, "dur": 3.345,
    "args": {
      "External id": 3049,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1116,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1116, "pid": 537, "tid": 537, "ts": 7429564839237.665,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839245.906, "dur": 3.788,
    "args": {
      "External id": 3050,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1117,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1117, "pid": 537, "tid": 537, "ts": 7429564839245.906,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839255.950, "dur": 4.860,
    "args": {
      "External id": 3051,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1118,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1118, "pid": 537, "tid": 537, "ts": 7429564839255.950,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839268.528, "dur": 4.777,
    "args": {
      "External id": 3053,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1119,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1119, "pid": 537, "tid": 537, "ts": 7429564839268.528,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839311.219, "dur": 0.213,
    "args": {
      "External id": 3061,      "cid": 370, "correlation": 1120
    }
  },
  {
    "ph": "f", "id": 1120, "pid": 537, "tid": 537, "ts": 7429564839311.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839314.336, "dur": 0.157,
    "args": {
      "External id": 3061,      "cid": 370, "correlation": 1121
    }
  },
  {
    "ph": "f", "id": 1121, "pid": 537, "tid": 537, "ts": 7429564839314.336,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839316.828, "dur": 0.178,
    "args": {
      "External id": 3061,      "cid": 370, "correlation": 1122
    }
  },
  {
    "ph": "f", "id": 1122, "pid": 537, "tid": 537, "ts": 7429564839316.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839318.952, "dur": 5.165,
    "args": {
      "External id": 3061,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1123,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1123, "pid": 537, "tid": 537, "ts": 7429564839318.952,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839394.379, "dur": 0.213,
    "args": {
      "External id": 3074,      "cid": 370, "correlation": 1124
    }
  },
  {
    "ph": "f", "id": 1124, "pid": 537, "tid": 537, "ts": 7429564839394.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839397.700, "dur": 0.155,
    "args": {
      "External id": 3074,      "cid": 370, "correlation": 1125
    }
  },
  {
    "ph": "f", "id": 1125, "pid": 537, "tid": 537, "ts": 7429564839397.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839400.142, "dur": 0.153,
    "args": {
      "External id": 3074,      "cid": 370, "correlation": 1126
    }
  },
  {
    "ph": "f", "id": 1126, "pid": 537, "tid": 537, "ts": 7429564839400.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839404.460, "dur": 6.234,
    "args": {
      "External id": 3074,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1127,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1127, "pid": 537, "tid": 537, "ts": 7429564839404.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839411.094, "dur": 3.187,
    "args": {
      "External id": 3074,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1128,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1128, "pid": 537, "tid": 537, "ts": 7429564839411.094,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839447.685, "dur": 0.197,
    "args": {
      "External id": 3087,      "cid": 370, "correlation": 1129
    }
  },
  {
    "ph": "f", "id": 1129, "pid": 537, "tid": 537, "ts": 7429564839447.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839450.901, "dur": 0.150,
    "args": {
      "External id": 3087,      "cid": 370, "correlation": 1130
    }
  },
  {
    "ph": "f", "id": 1130, "pid": 537, "tid": 537, "ts": 7429564839450.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839452.915, "dur": 0.141,
    "args": {
      "External id": 3087,      "cid": 370, "correlation": 1131
    }
  },
  {
    "ph": "f", "id": 1131, "pid": 537, "tid": 537, "ts": 7429564839452.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839456.476, "dur": 4.099,
    "args": {
      "External id": 3087,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1132,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1132, "pid": 537, "tid": 537, "ts": 7429564839456.476,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839460.846, "dur": 3.515,
    "args": {
      "External id": 3087,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1133,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1133, "pid": 537, "tid": 537, "ts": 7429564839460.846,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839487.521, "dur": 4.540,
    "args": {
      "External id": 3097,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1134,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1134, "pid": 537, "tid": 537, "ts": 7429564839487.521,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839514.133, "dur": 4.671,
    "args": {
      "External id": 3102,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1135,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1135, "pid": 537, "tid": 537, "ts": 7429564839514.133,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839532.703, "dur": 7.690,
    "args": {
      "External id": 3103,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1136,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1136, "pid": 537, "tid": 537, "ts": 7429564839532.703,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564839548.393, "dur": 0.270,
    "args": {
      "External id": 3104,      "cid": 290, "correlation": 1137
    }
  },
  {
    "ph": "f", "id": 1137, "pid": 537, "tid": 537, "ts": 7429564839548.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564839549.128, "dur": 41.732,
    "args": {
      "External id": 3104,            "bytes": 2097152, "cid": 108, "correlation": 1138, "ptr": "0x2"
    }
  },
  {
    "ph": "s", "id": 1138, "pid": 537, "tid": 537, "ts": 7429564839549.128,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839593.233, "dur": 4.688,
    "args": {
      "External id": 3104,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1139,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1139, "pid": 537, "tid": 537, "ts": 7429564839593.233,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839605.990, "dur": 5.404,
    "args": {
      "External id": 3105,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1140,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1140, "pid": 537, "tid": 537, "ts": 7429564839605.990,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839617.936, "dur": 3.537,
    "args": {
      "External id": 3106,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1141,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1141, "pid": 537, "tid": 537, "ts": 7429564839617.936,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839635.473, "dur": 4.603,
    "args": {
      "External id": 3111,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1142,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1142, "pid": 537, "tid": 537, "ts": 7429564839635.473,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839647.355, "dur": 5.178,
    "args": {
      "External id": 3112,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1143,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1143, "pid": 537, "tid": 537, "ts": 7429564839647.355,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839658.914, "dur": 4.771,
    "args": {
      "External id": 3113,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1144,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1144, "pid": 537, "tid": 537, "ts": 7429564839658.914,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839668.898, "dur": 3.505,
    "args": {
      "External id": 3114,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1145,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1145, "pid": 537, "tid": 537, "ts": 7429564839668.898,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839735.046, "dur": 4.439,
    "args": {
      "External id": 3131,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1146,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1146, "pid": 537, "tid": 537, "ts": 7429564839735.046,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839768.647, "dur": 4.543,
    "args": {
      "External id": 3149,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1147,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1147, "pid": 537, "tid": 537, "ts": 7429564839768.647,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839810.965, "dur": 0.326,
    "args": {
      "External id": 3162,      "cid": 370, "correlation": 1148
    }
  },
  {
    "ph": "f", "id": 1148, "pid": 537, "tid": 537, "ts": 7429564839810.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839815.450, "dur": 0.164,
    "args": {
      "External id": 3162,      "cid": 370, "correlation": 1149
    }
  },
  {
    "ph": "f", "id": 1149, "pid": 537, "tid": 537, "ts": 7429564839815.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839818.085, "dur": 0.190,
    "args": {
      "External id": 3162,      "cid": 370, "correlation": 1150
    }
  },
  {
    "ph": "f", "id": 1150, "pid": 537, "tid": 537, "ts": 7429564839818.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839821.234, "dur": 4.653,
    "args": {
      "External id": 3162,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1151,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1151, "pid": 537, "tid": 537, "ts": 7429564839821.234,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839838.122, "dur": 4.578,
    "args": {
      "External id": 3165,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1152,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1152, "pid": 537, "tid": 537, "ts": 7429564839838.122,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839856.709, "dur": 4.776,
    "args": {
      "External id": 3172,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1153,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1153, "pid": 537, "tid": 537, "ts": 7429564839856.709,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839873.109, "dur": 5.774,
    "args": {
      "External id": 3175,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1154,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1154, "pid": 537, "tid": 537, "ts": 7429564839873.109,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839921.355, "dur": 0.224,
    "args": {
      "External id": 3187,      "cid": 370, "correlation": 1155
    }
  },
  {
    "ph": "f", "id": 1155, "pid": 537, "tid": 537, "ts": 7429564839921.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839925.215, "dur": 0.146,
    "args": {
      "External id": 3187,      "cid": 370, "correlation": 1156
    }
  },
  {
    "ph": "f", "id": 1156, "pid": 537, "tid": 537, "ts": 7429564839925.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564839928.113, "dur": 0.190,
    "args": {
      "External id": 3187,      "cid": 370, "correlation": 1157
    }
  },
  {
    "ph": "f", "id": 1157, "pid": 537, "tid": 537, "ts": 7429564839928.113,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839930.431, "dur": 5.803,
    "args": {
      "External id": 3187,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1158,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1158, "pid": 537, "tid": 537, "ts": 7429564839930.431,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564839952.660, "dur": 3.550,
    "args": {
      "External id": 3196,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1159,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1159, "pid": 537, "tid": 537, "ts": 7429564839952.660,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840000.734, "dur": 0.286,
    "args": {
      "External id": 3206,      "cid": 370, "correlation": 1160
    }
  },
  {
    "ph": "f", "id": 1160, "pid": 537, "tid": 537, "ts": 7429564840000.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840003.912, "dur": 0.168,
    "args": {
      "External id": 3206,      "cid": 370, "correlation": 1161
    }
  },
  {
    "ph": "f", "id": 1161, "pid": 537, "tid": 537, "ts": 7429564840003.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840006.261, "dur": 0.191,
    "args": {
      "External id": 3206,      "cid": 370, "correlation": 1162
    }
  },
  {
    "ph": "f", "id": 1162, "pid": 537, "tid": 537, "ts": 7429564840006.261,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840008.145, "dur": 5.158,
    "args": {
      "External id": 3206,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1163,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1163, "pid": 537, "tid": 537, "ts": 7429564840008.145,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840026.211, "dur": 4.671,
    "args": {
      "External id": 3209,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1164,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1164, "pid": 537, "tid": 537, "ts": 7429564840026.211,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840049.312, "dur": 6.072,
    "args": {
      "External id": 3211,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1165,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1165, "pid": 537, "tid": 537, "ts": 7429564840049.312,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840066.088, "dur": 5.276,
    "args": {
      "External id": 3214,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1166,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1166, "pid": 537, "tid": 537, "ts": 7429564840066.088,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840080.206, "dur": 4.301,
    "args": {
      "External id": 3215,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1167,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1167, "pid": 537, "tid": 537, "ts": 7429564840080.206,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840090.423, "dur": 3.983,
    "args": {
      "External id": 3216,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1168,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1168, "pid": 537, "tid": 537, "ts": 7429564840090.423,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840100.828, "dur": 5.153,
    "args": {
      "External id": 3217,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1169,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1169, "pid": 537, "tid": 537, "ts": 7429564840100.828,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840115.189, "dur": 4.149,
    "args": {
      "External id": 3219,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1170,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1170, "pid": 537, "tid": 537, "ts": 7429564840115.189,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840155.861, "dur": 0.197,
    "args": {
      "External id": 3227,      "cid": 370, "correlation": 1171
    }
  },
  {
    "ph": "f", "id": 1171, "pid": 537, "tid": 537, "ts": 7429564840155.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840159.386, "dur": 0.142,
    "args": {
      "External id": 3227,      "cid": 370, "correlation": 1172
    }
  },
  {
    "ph": "f", "id": 1172, "pid": 537, "tid": 537, "ts": 7429564840159.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840161.567, "dur": 0.169,
    "args": {
      "External id": 3227,      "cid": 370, "correlation": 1173
    }
  },
  {
    "ph": "f", "id": 1173, "pid": 537, "tid": 537, "ts": 7429564840161.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840168.741, "dur": 6.502,
    "args": {
      "External id": 3227,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1174,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1174, "pid": 537, "tid": 537, "ts": 7429564840168.741,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840175.586, "dur": 4.084,
    "args": {
      "External id": 3227,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1175,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1175, "pid": 537, "tid": 537, "ts": 7429564840175.586,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564840196.607, "dur": 0.268,
    "args": {
      "External id": 3230,      "cid": 290, "correlation": 1176
    }
  },
  {
    "ph": "f", "id": 1176, "pid": 537, "tid": 537, "ts": 7429564840196.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564840197.266, "dur": 45.712,
    "args": {
      "External id": 3230,            "bytes": 2097152, "cid": 108, "correlation": 1177, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 1177, "pid": 537, "tid": 537, "ts": 7429564840197.266,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840246.272, "dur": 4.672,
    "args": {
      "External id": 3230,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1178,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1178, "pid": 537, "tid": 537, "ts": 7429564840246.272,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840281.820, "dur": 0.207,
    "args": {
      "External id": 3238,      "cid": 370, "correlation": 1179
    }
  },
  {
    "ph": "f", "id": 1179, "pid": 537, "tid": 537, "ts": 7429564840281.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840285.217, "dur": 0.142,
    "args": {
      "External id": 3238,      "cid": 370, "correlation": 1180
    }
  },
  {
    "ph": "f", "id": 1180, "pid": 537, "tid": 537, "ts": 7429564840285.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840287.387, "dur": 0.145,
    "args": {
      "External id": 3238,      "cid": 370, "correlation": 1181
    }
  },
  {
    "ph": "f", "id": 1181, "pid": 537, "tid": 537, "ts": 7429564840287.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840292.185, "dur": 5.563,
    "args": {
      "External id": 3238,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1182,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1182, "pid": 537, "tid": 537, "ts": 7429564840292.185,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840298.118, "dur": 3.688,
    "args": {
      "External id": 3238,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1183,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1183, "pid": 537, "tid": 537, "ts": 7429564840298.118,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564840311.111, "dur": 0.207,
    "args": {
      "External id": 3241,      "cid": 290, "correlation": 1184
    }
  },
  {
    "ph": "f", "id": 1184, "pid": 537, "tid": 537, "ts": 7429564840311.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564840311.614, "dur": 29.912,
    "args": {
      "External id": 3241,            "bytes": 2097152, "cid": 108, "correlation": 1185, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 1185, "pid": 537, "tid": 537, "ts": 7429564840311.614,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840344.043, "dur": 4.519,
    "args": {
      "External id": 3241,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1186,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1186, "pid": 537, "tid": 537, "ts": 7429564840344.043,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840378.166, "dur": 0.202,
    "args": {
      "External id": 3249,      "cid": 370, "correlation": 1187
    }
  },
  {
    "ph": "f", "id": 1187, "pid": 537, "tid": 537, "ts": 7429564840378.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840381.666, "dur": 0.138,
    "args": {
      "External id": 3249,      "cid": 370, "correlation": 1188
    }
  },
  {
    "ph": "f", "id": 1188, "pid": 537, "tid": 537, "ts": 7429564840381.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840384.018, "dur": 0.151,
    "args": {
      "External id": 3249,      "cid": 370, "correlation": 1189
    }
  },
  {
    "ph": "f", "id": 1189, "pid": 537, "tid": 537, "ts": 7429564840384.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840388.421, "dur": 5.822,
    "args": {
      "External id": 3249,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1190,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1190, "pid": 537, "tid": 537, "ts": 7429564840388.421,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840394.567, "dur": 3.694,
    "args": {
      "External id": 3249,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1191,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1191, "pid": 537, "tid": 537, "ts": 7429564840394.567,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840408.293, "dur": 5.028,
    "args": {
      "External id": 3252,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1192,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1192, "pid": 537, "tid": 537, "ts": 7429564840408.293,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840441.421, "dur": 4.096,
    "args": {
      "External id": 3254,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1193,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1193, "pid": 537, "tid": 537, "ts": 7429564840441.421,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840455.519, "dur": 7.369,
    "args": {
      "External id": 3257,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1194,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1194, "pid": 537, "tid": 537, "ts": 7429564840455.519,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840471.572, "dur": 3.752,
    "args": {
      "External id": 3258,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1195,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1195, "pid": 537, "tid": 537, "ts": 7429564840471.572,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840481.179, "dur": 6.056,
    "args": {
      "External id": 3259,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1196,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1196, "pid": 537, "tid": 537, "ts": 7429564840481.179,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840494.627, "dur": 3.577,
    "args": {
      "External id": 3260,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1197,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1197, "pid": 537, "tid": 537, "ts": 7429564840494.627,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840505.396, "dur": 5.121,
    "args": {
      "External id": 3262,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1198,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1198, "pid": 537, "tid": 537, "ts": 7429564840505.396,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840550.642, "dur": 0.283,
    "args": {
      "External id": 3270,      "cid": 370, "correlation": 1199
    }
  },
  {
    "ph": "f", "id": 1199, "pid": 537, "tid": 537, "ts": 7429564840550.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840553.620, "dur": 0.156,
    "args": {
      "External id": 3270,      "cid": 370, "correlation": 1200
    }
  },
  {
    "ph": "f", "id": 1200, "pid": 537, "tid": 537, "ts": 7429564840553.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840555.684, "dur": 0.168,
    "args": {
      "External id": 3270,      "cid": 370, "correlation": 1201
    }
  },
  {
    "ph": "f", "id": 1201, "pid": 537, "tid": 537, "ts": 7429564840555.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840557.941, "dur": 4.680,
    "args": {
      "External id": 3270,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1202,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1202, "pid": 537, "tid": 537, "ts": 7429564840557.941,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840602.370, "dur": 0.182,
    "args": {
      "External id": 3283,      "cid": 370, "correlation": 1203
    }
  },
  {
    "ph": "f", "id": 1203, "pid": 537, "tid": 537, "ts": 7429564840602.370,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840605.821, "dur": 0.170,
    "args": {
      "External id": 3283,      "cid": 370, "correlation": 1204
    }
  },
  {
    "ph": "f", "id": 1204, "pid": 537, "tid": 537, "ts": 7429564840605.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840607.646, "dur": 0.140,
    "args": {
      "External id": 3283,      "cid": 370, "correlation": 1205
    }
  },
  {
    "ph": "f", "id": 1205, "pid": 537, "tid": 537, "ts": 7429564840607.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840611.854, "dur": 4.643,
    "args": {
      "External id": 3283,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1206,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1206, "pid": 537, "tid": 537, "ts": 7429564840611.854,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840616.858, "dur": 4.833,
    "args": {
      "External id": 3283,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1207,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1207, "pid": 537, "tid": 537, "ts": 7429564840616.858,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840655.506, "dur": 0.203,
    "args": {
      "External id": 3296,      "cid": 370, "correlation": 1208
    }
  },
  {
    "ph": "f", "id": 1208, "pid": 537, "tid": 537, "ts": 7429564840655.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840658.217, "dur": 0.153,
    "args": {
      "External id": 3296,      "cid": 370, "correlation": 1209
    }
  },
  {
    "ph": "f", "id": 1209, "pid": 537, "tid": 537, "ts": 7429564840658.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840660.416, "dur": 0.147,
    "args": {
      "External id": 3296,      "cid": 370, "correlation": 1210
    }
  },
  {
    "ph": "f", "id": 1210, "pid": 537, "tid": 537, "ts": 7429564840660.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840663.948, "dur": 5.534,
    "args": {
      "External id": 3296,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1211,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1211, "pid": 537, "tid": 537, "ts": 7429564840663.948,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840669.954, "dur": 3.488,
    "args": {
      "External id": 3296,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1212,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1212, "pid": 537, "tid": 537, "ts": 7429564840669.954,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840697.192, "dur": 4.271,
    "args": {
      "External id": 3306,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1213,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1213, "pid": 537, "tid": 537, "ts": 7429564840697.192,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840726.592, "dur": 5.160,
    "args": {
      "External id": 3311,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1214,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1214, "pid": 537, "tid": 537, "ts": 7429564840726.592,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840744.559, "dur": 7.170,
    "args": {
      "External id": 3312,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1215,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1215, "pid": 537, "tid": 537, "ts": 7429564840744.559,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840759.839, "dur": 4.619,
    "args": {
      "External id": 3313,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1216,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1216, "pid": 537, "tid": 537, "ts": 7429564840759.839,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840770.199, "dur": 5.051,
    "args": {
      "External id": 3314,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1217,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1217, "pid": 537, "tid": 537, "ts": 7429564840770.199,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840781.181, "dur": 3.226,
    "args": {
      "External id": 3315,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1218,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1218, "pid": 537, "tid": 537, "ts": 7429564840781.181,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840798.837, "dur": 3.533,
    "args": {
      "External id": 3320,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1219,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1219, "pid": 537, "tid": 537, "ts": 7429564840798.837,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840809.982, "dur": 6.338,
    "args": {
      "External id": 3321,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1220,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1220, "pid": 537, "tid": 537, "ts": 7429564840809.982,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840822.856, "dur": 3.232,
    "args": {
      "External id": 3322,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1221,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1221, "pid": 537, "tid": 537, "ts": 7429564840822.856,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840830.486, "dur": 3.391,
    "args": {
      "External id": 3323,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1222,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1222, "pid": 537, "tid": 537, "ts": 7429564840830.486,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840890.313, "dur": 5.788,
    "args": {
      "External id": 3340,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1223,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1223, "pid": 537, "tid": 537, "ts": 7429564840890.313,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840926.137, "dur": 3.907,
    "args": {
      "External id": 3358,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1224,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1224, "pid": 537, "tid": 537, "ts": 7429564840926.137,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840964.159, "dur": 0.273,
    "args": {
      "External id": 3371,      "cid": 370, "correlation": 1225
    }
  },
  {
    "ph": "f", "id": 1225, "pid": 537, "tid": 537, "ts": 7429564840964.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840968.057, "dur": 0.161,
    "args": {
      "External id": 3371,      "cid": 370, "correlation": 1226
    }
  },
  {
    "ph": "f", "id": 1226, "pid": 537, "tid": 537, "ts": 7429564840968.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564840970.263, "dur": 0.173,
    "args": {
      "External id": 3371,      "cid": 370, "correlation": 1227
    }
  },
  {
    "ph": "f", "id": 1227, "pid": 537, "tid": 537, "ts": 7429564840970.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840972.797, "dur": 3.914,
    "args": {
      "External id": 3371,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1228,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1228, "pid": 537, "tid": 537, "ts": 7429564840972.797,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564840987.637, "dur": 4.830,
    "args": {
      "External id": 3374,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1229,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1229, "pid": 537, "tid": 537, "ts": 7429564840987.637,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841006.859, "dur": 4.174,
    "args": {
      "External id": 3381,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1230,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1230, "pid": 537, "tid": 537, "ts": 7429564841006.859,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841022.875, "dur": 5.175,
    "args": {
      "External id": 3384,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1231,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1231, "pid": 537, "tid": 537, "ts": 7429564841022.875,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841070.714, "dur": 0.232,
    "args": {
      "External id": 3396,      "cid": 370, "correlation": 1232
    }
  },
  {
    "ph": "f", "id": 1232, "pid": 537, "tid": 537, "ts": 7429564841070.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841074.653, "dur": 0.144,
    "args": {
      "External id": 3396,      "cid": 370, "correlation": 1233
    }
  },
  {
    "ph": "f", "id": 1233, "pid": 537, "tid": 537, "ts": 7429564841074.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841077.619, "dur": 0.142,
    "args": {
      "External id": 3396,      "cid": 370, "correlation": 1234
    }
  },
  {
    "ph": "f", "id": 1234, "pid": 537, "tid": 537, "ts": 7429564841077.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841079.644, "dur": 5.993,
    "args": {
      "External id": 3396,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1235,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1235, "pid": 537, "tid": 537, "ts": 7429564841079.644,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841102.882, "dur": 3.998,
    "args": {
      "External id": 3405,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1236,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1236, "pid": 537, "tid": 537, "ts": 7429564841102.882,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841148.407, "dur": 0.255,
    "args": {
      "External id": 3415,      "cid": 370, "correlation": 1237
    }
  },
  {
    "ph": "f", "id": 1237, "pid": 537, "tid": 537, "ts": 7429564841148.407,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841151.999, "dur": 0.156,
    "args": {
      "External id": 3415,      "cid": 370, "correlation": 1238
    }
  },
  {
    "ph": "f", "id": 1238, "pid": 537, "tid": 537, "ts": 7429564841151.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841153.985, "dur": 0.190,
    "args": {
      "External id": 3415,      "cid": 370, "correlation": 1239
    }
  },
  {
    "ph": "f", "id": 1239, "pid": 537, "tid": 537, "ts": 7429564841153.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841155.824, "dur": 4.505,
    "args": {
      "External id": 3415,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1240,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1240, "pid": 537, "tid": 537, "ts": 7429564841155.824,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841172.484, "dur": 5.145,
    "args": {
      "External id": 3418,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1241,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1241, "pid": 537, "tid": 537, "ts": 7429564841172.484,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841195.082, "dur": 5.210,
    "args": {
      "External id": 3420,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1242,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1242, "pid": 537, "tid": 537, "ts": 7429564841195.082,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841209.680, "dur": 4.392,
    "args": {
      "External id": 3423,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1243,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1243, "pid": 537, "tid": 537, "ts": 7429564841209.680,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841223.396, "dur": 4.985,
    "args": {
      "External id": 3424,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1244,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1244, "pid": 537, "tid": 537, "ts": 7429564841223.396,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841234.574, "dur": 4.137,
    "args": {
      "External id": 3425,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1245,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1245, "pid": 537, "tid": 537, "ts": 7429564841234.574,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841245.881, "dur": 3.561,
    "args": {
      "External id": 3426,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1246,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1246, "pid": 537, "tid": 537, "ts": 7429564841245.881,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841257.085, "dur": 5.092,
    "args": {
      "External id": 3428,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1247,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1247, "pid": 537, "tid": 537, "ts": 7429564841257.085,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564841288.702, "dur": 0.282,
    "args": {
      "External id": 3436,      "cid": 290, "correlation": 1248
    }
  },
  {
    "ph": "f", "id": 1248, "pid": 537, "tid": 537, "ts": 7429564841288.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564841289.395, "dur": 53.760,
    "args": {
      "External id": 3436,            "bytes": 2097152, "cid": 108, "correlation": 1249, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1249, "pid": 537, "tid": 537, "ts": 7429564841289.395,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841355.806, "dur": 0.276,
    "args": {
      "External id": 3436,      "cid": 370, "correlation": 1250
    }
  },
  {
    "ph": "f", "id": 1250, "pid": 537, "tid": 537, "ts": 7429564841355.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841359.786, "dur": 0.139,
    "args": {
      "External id": 3436,      "cid": 370, "correlation": 1251
    }
  },
  {
    "ph": "f", "id": 1251, "pid": 537, "tid": 537, "ts": 7429564841359.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841362.396, "dur": 0.170,
    "args": {
      "External id": 3436,      "cid": 370, "correlation": 1252
    }
  },
  {
    "ph": "f", "id": 1252, "pid": 537, "tid": 537, "ts": 7429564841362.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841371.108, "dur": 6.617,
    "args": {
      "External id": 3436,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1253,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1253, "pid": 537, "tid": 537, "ts": 7429564841371.108,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841378.260, "dur": 4.391,
    "args": {
      "External id": 3436,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1254,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1254, "pid": 537, "tid": 537, "ts": 7429564841378.260,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841402.747, "dur": 6.007,
    "args": {
      "External id": 3439,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1255,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1255, "pid": 537, "tid": 537, "ts": 7429564841402.747,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564841429.834, "dur": 0.280,
    "args": {
      "External id": 3447,      "cid": 290, "correlation": 1256
    }
  },
  {
    "ph": "f", "id": 1256, "pid": 537, "tid": 537, "ts": 7429564841429.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564841430.443, "dur": 30.282,
    "args": {
      "External id": 3447,            "bytes": 2097152, "cid": 108, "correlation": 1257, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1257, "pid": 537, "tid": 537, "ts": 7429564841430.443,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841471.909, "dur": 0.204,
    "args": {
      "External id": 3447,      "cid": 370, "correlation": 1258
    }
  },
  {
    "ph": "f", "id": 1258, "pid": 537, "tid": 537, "ts": 7429564841471.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841475.114, "dur": 0.156,
    "args": {
      "External id": 3447,      "cid": 370, "correlation": 1259
    }
  },
  {
    "ph": "f", "id": 1259, "pid": 537, "tid": 537, "ts": 7429564841475.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841477.629, "dur": 0.163,
    "args": {
      "External id": 3447,      "cid": 370, "correlation": 1260
    }
  },
  {
    "ph": "f", "id": 1260, "pid": 537, "tid": 537, "ts": 7429564841477.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841482.243, "dur": 5.088,
    "args": {
      "External id": 3447,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1261,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1261, "pid": 537, "tid": 537, "ts": 7429564841482.243,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841487.843, "dur": 5.018,
    "args": {
      "External id": 3447,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1262,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1262, "pid": 537, "tid": 537, "ts": 7429564841487.843,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841503.508, "dur": 4.569,
    "args": {
      "External id": 3450,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1263,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1263, "pid": 537, "tid": 537, "ts": 7429564841503.508,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841533.111, "dur": 0.172,
    "args": {
      "External id": 3458,      "cid": 370, "correlation": 1264
    }
  },
  {
    "ph": "f", "id": 1264, "pid": 537, "tid": 537, "ts": 7429564841533.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841536.175, "dur": 0.156,
    "args": {
      "External id": 3458,      "cid": 370, "correlation": 1265
    }
  },
  {
    "ph": "f", "id": 1265, "pid": 537, "tid": 537, "ts": 7429564841536.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841538.307, "dur": 0.150,
    "args": {
      "External id": 3458,      "cid": 370, "correlation": 1266
    }
  },
  {
    "ph": "f", "id": 1266, "pid": 537, "tid": 537, "ts": 7429564841538.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841542.078, "dur": 4.596,
    "args": {
      "External id": 3458,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1267,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1267, "pid": 537, "tid": 537, "ts": 7429564841542.078,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841547.159, "dur": 3.580,
    "args": {
      "External id": 3458,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1268,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1268, "pid": 537, "tid": 537, "ts": 7429564841547.159,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841560.979, "dur": 5.184,
    "args": {
      "External id": 3461,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1269,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1269, "pid": 537, "tid": 537, "ts": 7429564841560.979,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841593.918, "dur": 4.731,
    "args": {
      "External id": 3463,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1270,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1270, "pid": 537, "tid": 537, "ts": 7429564841593.918,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841607.684, "dur": 4.742,
    "args": {
      "External id": 3466,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1271,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1271, "pid": 537, "tid": 537, "ts": 7429564841607.684,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841620.344, "dur": 4.119,
    "args": {
      "External id": 3467,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1272,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1272, "pid": 537, "tid": 537, "ts": 7429564841620.344,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841630.915, "dur": 4.048,
    "args": {
      "External id": 3468,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1273,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1273, "pid": 537, "tid": 537, "ts": 7429564841630.915,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841641.119, "dur": 3.634,
    "args": {
      "External id": 3469,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1274,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1274, "pid": 537, "tid": 537, "ts": 7429564841641.119,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841653.120, "dur": 5.008,
    "args": {
      "External id": 3471,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1275,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1275, "pid": 537, "tid": 537, "ts": 7429564841653.120,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841697.029, "dur": 0.192,
    "args": {
      "External id": 3479,      "cid": 370, "correlation": 1276
    }
  },
  {
    "ph": "f", "id": 1276, "pid": 537, "tid": 537, "ts": 7429564841697.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841700.818, "dur": 0.161,
    "args": {
      "External id": 3479,      "cid": 370, "correlation": 1277
    }
  },
  {
    "ph": "f", "id": 1277, "pid": 537, "tid": 537, "ts": 7429564841700.818,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841703.423, "dur": 0.162,
    "args": {
      "External id": 3479,      "cid": 370, "correlation": 1278
    }
  },
  {
    "ph": "f", "id": 1278, "pid": 537, "tid": 537, "ts": 7429564841703.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841705.507, "dur": 4.900,
    "args": {
      "External id": 3479,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1279,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1279, "pid": 537, "tid": 537, "ts": 7429564841705.507,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841746.324, "dur": 0.253,
    "args": {
      "External id": 3492,      "cid": 370, "correlation": 1280
    }
  },
  {
    "ph": "f", "id": 1280, "pid": 537, "tid": 537, "ts": 7429564841746.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841749.465, "dur": 0.155,
    "args": {
      "External id": 3492,      "cid": 370, "correlation": 1281
    }
  },
  {
    "ph": "f", "id": 1281, "pid": 537, "tid": 537, "ts": 7429564841749.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841751.541, "dur": 0.166,
    "args": {
      "External id": 3492,      "cid": 370, "correlation": 1282
    }
  },
  {
    "ph": "f", "id": 1282, "pid": 537, "tid": 537, "ts": 7429564841751.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841755.699, "dur": 5.836,
    "args": {
      "External id": 3492,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1283,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1283, "pid": 537, "tid": 537, "ts": 7429564841755.699,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841762.060, "dur": 3.603,
    "args": {
      "External id": 3492,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1284,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1284, "pid": 537, "tid": 537, "ts": 7429564841762.060,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841797.778, "dur": 0.218,
    "args": {
      "External id": 3505,      "cid": 370, "correlation": 1285
    }
  },
  {
    "ph": "f", "id": 1285, "pid": 537, "tid": 537, "ts": 7429564841797.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841800.675, "dur": 0.151,
    "args": {
      "External id": 3505,      "cid": 370, "correlation": 1286
    }
  },
  {
    "ph": "f", "id": 1286, "pid": 537, "tid": 537, "ts": 7429564841800.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564841802.776, "dur": 0.158,
    "args": {
      "External id": 3505,      "cid": 370, "correlation": 1287
    }
  },
  {
    "ph": "f", "id": 1287, "pid": 537, "tid": 537, "ts": 7429564841802.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841806.139, "dur": 4.222,
    "args": {
      "External id": 3505,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1288,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1288, "pid": 537, "tid": 537, "ts": 7429564841806.139,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841810.666, "dur": 3.474,
    "args": {
      "External id": 3505,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1289,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1289, "pid": 537, "tid": 537, "ts": 7429564841810.666,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564841836.489, "dur": 0.229,
    "args": {
      "External id": 3515,      "cid": 290, "correlation": 1290
    }
  },
  {
    "ph": "f", "id": 1290, "pid": 537, "tid": 537, "ts": 7429564841836.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564841837.099, "dur": 32.683,
    "args": {
      "External id": 3515,            "bytes": 2097152, "cid": 108, "correlation": 1291, "ptr": "0x7fee32880142"
    }
  },
  {
    "ph": "s", "id": 1291, "pid": 537, "tid": 537, "ts": 7429564841837.099,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841873.092, "dur": 5.149,
    "args": {
      "External id": 3515,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1292,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1292, "pid": 537, "tid": 537, "ts": 7429564841873.092,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841901.307, "dur": 6.723,
    "args": {
      "External id": 3520,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1293,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1293, "pid": 537, "tid": 537, "ts": 7429564841901.307,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841924.514, "dur": 6.103,
    "args": {
      "External id": 3521,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1294,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1294, "pid": 537, "tid": 537, "ts": 7429564841924.514,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841938.935, "dur": 4.567,
    "args": {
      "External id": 3522,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1295,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1295, "pid": 537, "tid": 537, "ts": 7429564841938.935,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841949.079, "dur": 3.844,
    "args": {
      "External id": 3523,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1296,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1296, "pid": 537, "tid": 537, "ts": 7429564841949.079,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841958.652, "dur": 3.798,
    "args": {
      "External id": 3524,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1297,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1297, "pid": 537, "tid": 537, "ts": 7429564841958.652,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841976.265, "dur": 4.496,
    "args": {
      "External id": 3529,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1298,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1298, "pid": 537, "tid": 537, "ts": 7429564841976.265,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564841988.460, "dur": 5.924,
    "args": {
      "External id": 3530,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1299,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1299, "pid": 537, "tid": 537, "ts": 7429564841988.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842000.475, "dur": 3.283,
    "args": {
      "External id": 3531,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1300,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1300, "pid": 537, "tid": 537, "ts": 7429564842000.475,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842008.112, "dur": 3.667,
    "args": {
      "External id": 3532,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1301,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1301, "pid": 537, "tid": 537, "ts": 7429564842008.112,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842093.787, "dur": 5.940,
    "args": {
      "External id": 3549,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1302,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1302, "pid": 537, "tid": 537, "ts": 7429564842093.787,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842136.580, "dur": 3.974,
    "args": {
      "External id": 3567,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1303,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1303, "pid": 537, "tid": 537, "ts": 7429564842136.580,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842180.292, "dur": 0.334,
    "args": {
      "External id": 3580,      "cid": 370, "correlation": 1304
    }
  },
  {
    "ph": "f", "id": 1304, "pid": 537, "tid": 537, "ts": 7429564842180.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842186.025, "dur": 0.170,
    "args": {
      "External id": 3580,      "cid": 370, "correlation": 1305
    }
  },
  {
    "ph": "f", "id": 1305, "pid": 537, "tid": 537, "ts": 7429564842186.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842189.146, "dur": 0.182,
    "args": {
      "External id": 3580,      "cid": 370, "correlation": 1306
    }
  },
  {
    "ph": "f", "id": 1306, "pid": 537, "tid": 537, "ts": 7429564842189.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842193.091, "dur": 6.157,
    "args": {
      "External id": 3580,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1307,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1307, "pid": 537, "tid": 537, "ts": 7429564842193.091,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842212.988, "dur": 4.542,
    "args": {
      "External id": 3583,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1308,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1308, "pid": 537, "tid": 537, "ts": 7429564842212.988,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842270.531, "dur": 4.559,
    "args": {
      "External id": 3590,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1309,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1309, "pid": 537, "tid": 537, "ts": 7429564842270.531,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842286.880, "dur": 6.690,
    "args": {
      "External id": 3593,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1310,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1310, "pid": 537, "tid": 537, "ts": 7429564842286.880,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842334.332, "dur": 0.169,
    "args": {
      "External id": 3605,      "cid": 370, "correlation": 1311
    }
  },
  {
    "ph": "f", "id": 1311, "pid": 537, "tid": 537, "ts": 7429564842334.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842337.926, "dur": 0.162,
    "args": {
      "External id": 3605,      "cid": 370, "correlation": 1312
    }
  },
  {
    "ph": "f", "id": 1312, "pid": 537, "tid": 537, "ts": 7429564842337.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842340.941, "dur": 0.175,
    "args": {
      "External id": 3605,      "cid": 370, "correlation": 1313
    }
  },
  {
    "ph": "f", "id": 1313, "pid": 537, "tid": 537, "ts": 7429564842340.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842343.228, "dur": 6.051,
    "args": {
      "External id": 3605,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1314,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1314, "pid": 537, "tid": 537, "ts": 7429564842343.228,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842365.254, "dur": 3.482,
    "args": {
      "External id": 3614,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1315,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1315, "pid": 537, "tid": 537, "ts": 7429564842365.254,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842412.495, "dur": 0.212,
    "args": {
      "External id": 3624,      "cid": 370, "correlation": 1316
    }
  },
  {
    "ph": "f", "id": 1316, "pid": 537, "tid": 537, "ts": 7429564842412.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842415.765, "dur": 0.171,
    "args": {
      "External id": 3624,      "cid": 370, "correlation": 1317
    }
  },
  {
    "ph": "f", "id": 1317, "pid": 537, "tid": 537, "ts": 7429564842415.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842418.108, "dur": 0.307,
    "args": {
      "External id": 3624,      "cid": 370, "correlation": 1318
    }
  },
  {
    "ph": "f", "id": 1318, "pid": 537, "tid": 537, "ts": 7429564842418.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842420.068, "dur": 5.948,
    "args": {
      "External id": 3624,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1319,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1319, "pid": 537, "tid": 537, "ts": 7429564842420.068,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842438.441, "dur": 4.414,
    "args": {
      "External id": 3627,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1320,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1320, "pid": 537, "tid": 537, "ts": 7429564842438.441,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842460.780, "dur": 3.850,
    "args": {
      "External id": 3629,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1321,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1321, "pid": 537, "tid": 537, "ts": 7429564842460.780,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842475.236, "dur": 5.055,
    "args": {
      "External id": 3632,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1322,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1322, "pid": 537, "tid": 537, "ts": 7429564842475.236,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842488.506, "dur": 3.979,
    "args": {
      "External id": 3633,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1323,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1323, "pid": 537, "tid": 537, "ts": 7429564842488.506,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842499.074, "dur": 4.815,
    "args": {
      "External id": 3634,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1324,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1324, "pid": 537, "tid": 537, "ts": 7429564842499.074,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842510.864, "dur": 4.820,
    "args": {
      "External id": 3635,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1325,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1325, "pid": 537, "tid": 537, "ts": 7429564842510.864,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842523.302, "dur": 4.777,
    "args": {
      "External id": 3637,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1326,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1326, "pid": 537, "tid": 537, "ts": 7429564842523.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842563.469, "dur": 0.188,
    "args": {
      "External id": 3645,      "cid": 370, "correlation": 1327
    }
  },
  {
    "ph": "f", "id": 1327, "pid": 537, "tid": 537, "ts": 7429564842563.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842566.841, "dur": 0.147,
    "args": {
      "External id": 3645,      "cid": 370, "correlation": 1328
    }
  },
  {
    "ph": "f", "id": 1328, "pid": 537, "tid": 537, "ts": 7429564842566.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842568.966, "dur": 0.161,
    "args": {
      "External id": 3645,      "cid": 370, "correlation": 1329
    }
  },
  {
    "ph": "f", "id": 1329, "pid": 537, "tid": 537, "ts": 7429564842568.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842576.841, "dur": 5.352,
    "args": {
      "External id": 3645,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1330,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1330, "pid": 537, "tid": 537, "ts": 7429564842576.841,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842582.541, "dur": 3.709,
    "args": {
      "External id": 3645,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1331,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1331, "pid": 537, "tid": 537, "ts": 7429564842582.541,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564842603.126, "dur": 0.313,
    "args": {
      "External id": 3648,      "cid": 290, "correlation": 1332
    }
  },
  {
    "ph": "f", "id": 1332, "pid": 537, "tid": 537, "ts": 7429564842603.126,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564842603.909, "dur": 55.967,
    "args": {
      "External id": 3648,            "bytes": 2097152, "cid": 108, "correlation": 1333, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 1333, "pid": 537, "tid": 537, "ts": 7429564842603.909,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842663.415, "dur": 4.926,
    "args": {
      "External id": 3648,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1334,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1334, "pid": 537, "tid": 537, "ts": 7429564842663.415,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842698.685, "dur": 0.199,
    "args": {
      "External id": 3656,      "cid": 370, "correlation": 1335
    }
  },
  {
    "ph": "f", "id": 1335, "pid": 537, "tid": 537, "ts": 7429564842698.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842702.047, "dur": 0.154,
    "args": {
      "External id": 3656,      "cid": 370, "correlation": 1336
    }
  },
  {
    "ph": "f", "id": 1336, "pid": 537, "tid": 537, "ts": 7429564842702.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842704.295, "dur": 0.153,
    "args": {
      "External id": 3656,      "cid": 370, "correlation": 1337
    }
  },
  {
    "ph": "f", "id": 1337, "pid": 537, "tid": 537, "ts": 7429564842704.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842709.435, "dur": 5.599,
    "args": {
      "External id": 3656,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1338,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1338, "pid": 537, "tid": 537, "ts": 7429564842709.435,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842715.351, "dur": 4.497,
    "args": {
      "External id": 3656,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1339,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1339, "pid": 537, "tid": 537, "ts": 7429564842715.351,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564842728.399, "dur": 0.167,
    "args": {
      "External id": 3659,      "cid": 290, "correlation": 1340
    }
  },
  {
    "ph": "f", "id": 1340, "pid": 537, "tid": 537, "ts": 7429564842728.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564842728.874, "dur": 30.890,
    "args": {
      "External id": 3659,            "bytes": 2097152, "cid": 108, "correlation": 1341, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 1341, "pid": 537, "tid": 537, "ts": 7429564842728.874,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842762.273, "dur": 4.567,
    "args": {
      "External id": 3659,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1342,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1342, "pid": 537, "tid": 537, "ts": 7429564842762.273,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842793.351, "dur": 0.231,
    "args": {
      "External id": 3667,      "cid": 370, "correlation": 1343
    }
  },
  {
    "ph": "f", "id": 1343, "pid": 537, "tid": 537, "ts": 7429564842793.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842796.744, "dur": 0.153,
    "args": {
      "External id": 3667,      "cid": 370, "correlation": 1344
    }
  },
  {
    "ph": "f", "id": 1344, "pid": 537, "tid": 537, "ts": 7429564842796.744,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842798.866, "dur": 0.142,
    "args": {
      "External id": 3667,      "cid": 370, "correlation": 1345
    }
  },
  {
    "ph": "f", "id": 1345, "pid": 537, "tid": 537, "ts": 7429564842798.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842802.969, "dur": 4.937,
    "args": {
      "External id": 3667,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1346,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1346, "pid": 537, "tid": 537, "ts": 7429564842802.969,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842808.203, "dur": 3.397,
    "args": {
      "External id": 3667,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1347,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1347, "pid": 537, "tid": 537, "ts": 7429564842808.203,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842822.005, "dur": 4.799,
    "args": {
      "External id": 3670,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1348,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1348, "pid": 537, "tid": 537, "ts": 7429564842822.005,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842855.549, "dur": 5.513,
    "args": {
      "External id": 3672,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1349,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1349, "pid": 537, "tid": 537, "ts": 7429564842855.549,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842870.680, "dur": 5.338,
    "args": {
      "External id": 3675,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1350,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1350, "pid": 537, "tid": 537, "ts": 7429564842870.680,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842884.604, "dur": 3.762,
    "args": {
      "External id": 3676,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1351,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1351, "pid": 537, "tid": 537, "ts": 7429564842884.604,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842894.562, "dur": 4.355,
    "args": {
      "External id": 3677,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1352,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1352, "pid": 537, "tid": 537, "ts": 7429564842894.562,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842904.719, "dur": 3.321,
    "args": {
      "External id": 3678,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1353,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1353, "pid": 537, "tid": 537, "ts": 7429564842904.719,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842916.572, "dur": 3.621,
    "args": {
      "External id": 3680,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1354,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1354, "pid": 537, "tid": 537, "ts": 7429564842916.572,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842958.074, "dur": 0.300,
    "args": {
      "External id": 3688,      "cid": 370, "correlation": 1355
    }
  },
  {
    "ph": "f", "id": 1355, "pid": 537, "tid": 537, "ts": 7429564842958.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842961.595, "dur": 0.173,
    "args": {
      "External id": 3688,      "cid": 370, "correlation": 1356
    }
  },
  {
    "ph": "f", "id": 1356, "pid": 537, "tid": 537, "ts": 7429564842961.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564842963.995, "dur": 0.173,
    "args": {
      "External id": 3688,      "cid": 370, "correlation": 1357
    }
  },
  {
    "ph": "f", "id": 1357, "pid": 537, "tid": 537, "ts": 7429564842963.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564842966.249, "dur": 4.652,
    "args": {
      "External id": 3688,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1358,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1358, "pid": 537, "tid": 537, "ts": 7429564842966.249,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843005.965, "dur": 0.175,
    "args": {
      "External id": 3701,      "cid": 370, "correlation": 1359
    }
  },
  {
    "ph": "f", "id": 1359, "pid": 537, "tid": 537, "ts": 7429564843005.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843009.335, "dur": 0.155,
    "args": {
      "External id": 3701,      "cid": 370, "correlation": 1360
    }
  },
  {
    "ph": "f", "id": 1360, "pid": 537, "tid": 537, "ts": 7429564843009.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843011.416, "dur": 0.164,
    "args": {
      "External id": 3701,      "cid": 370, "correlation": 1361
    }
  },
  {
    "ph": "f", "id": 1361, "pid": 537, "tid": 537, "ts": 7429564843011.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843015.997, "dur": 6.027,
    "args": {
      "External id": 3701,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1362,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1362, "pid": 537, "tid": 537, "ts": 7429564843015.997,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843022.417, "dur": 4.130,
    "args": {
      "External id": 3701,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1363,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1363, "pid": 537, "tid": 537, "ts": 7429564843022.417,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843059.994, "dur": 0.188,
    "args": {
      "External id": 3714,      "cid": 370, "correlation": 1364
    }
  },
  {
    "ph": "f", "id": 1364, "pid": 537, "tid": 537, "ts": 7429564843059.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843062.643, "dur": 0.145,
    "args": {
      "External id": 3714,      "cid": 370, "correlation": 1365
    }
  },
  {
    "ph": "f", "id": 1365, "pid": 537, "tid": 537, "ts": 7429564843062.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843064.624, "dur": 0.155,
    "args": {
      "External id": 3714,      "cid": 370, "correlation": 1366
    }
  },
  {
    "ph": "f", "id": 1366, "pid": 537, "tid": 537, "ts": 7429564843064.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843068.006, "dur": 4.478,
    "args": {
      "External id": 3714,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1367,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1367, "pid": 537, "tid": 537, "ts": 7429564843068.006,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843072.771, "dur": 4.820,
    "args": {
      "External id": 3714,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1368,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1368, "pid": 537, "tid": 537, "ts": 7429564843072.771,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843101.586, "dur": 4.594,
    "args": {
      "External id": 3724,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1369,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1369, "pid": 537, "tid": 537, "ts": 7429564843101.586,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843130.099, "dur": 4.247,
    "args": {
      "External id": 3729,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1370,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1370, "pid": 537, "tid": 537, "ts": 7429564843130.099,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843148.853, "dur": 7.220,
    "args": {
      "External id": 3730,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1371,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1371, "pid": 537, "tid": 537, "ts": 7429564843148.853,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843163.836, "dur": 3.773,
    "args": {
      "External id": 3731,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1372,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1372, "pid": 537, "tid": 537, "ts": 7429564843163.836,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843173.075, "dur": 3.593,
    "args": {
      "External id": 3732,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1373,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1373, "pid": 537, "tid": 537, "ts": 7429564843173.075,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843182.449, "dur": 3.900,
    "args": {
      "External id": 3733,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1374,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1374, "pid": 537, "tid": 537, "ts": 7429564843182.449,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843200.168, "dur": 3.839,
    "args": {
      "External id": 3738,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1375,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1375, "pid": 537, "tid": 537, "ts": 7429564843200.168,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843211.558, "dur": 5.428,
    "args": {
      "External id": 3739,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1376,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1376, "pid": 537, "tid": 537, "ts": 7429564843211.558,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843223.343, "dur": 4.949,
    "args": {
      "External id": 3740,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1377,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1377, "pid": 537, "tid": 537, "ts": 7429564843223.343,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843233.791, "dur": 3.602,
    "args": {
      "External id": 3741,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1378,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1378, "pid": 537, "tid": 537, "ts": 7429564843233.791,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843297.211, "dur": 4.348,
    "args": {
      "External id": 3758,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1379,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1379, "pid": 537, "tid": 537, "ts": 7429564843297.211,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843332.444, "dur": 4.738,
    "args": {
      "External id": 3776,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1380,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1380, "pid": 537, "tid": 537, "ts": 7429564843332.444,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843375.101, "dur": 0.255,
    "args": {
      "External id": 3789,      "cid": 370, "correlation": 1381
    }
  },
  {
    "ph": "f", "id": 1381, "pid": 537, "tid": 537, "ts": 7429564843375.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843378.998, "dur": 0.160,
    "args": {
      "External id": 3789,      "cid": 370, "correlation": 1382
    }
  },
  {
    "ph": "f", "id": 1382, "pid": 537, "tid": 537, "ts": 7429564843378.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843381.463, "dur": 0.192,
    "args": {
      "External id": 3789,      "cid": 370, "correlation": 1383
    }
  },
  {
    "ph": "f", "id": 1383, "pid": 537, "tid": 537, "ts": 7429564843381.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843383.840, "dur": 5.843,
    "args": {
      "External id": 3789,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1384,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1384, "pid": 537, "tid": 537, "ts": 7429564843383.840,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843401.086, "dur": 4.020,
    "args": {
      "External id": 3792,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1385,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1385, "pid": 537, "tid": 537, "ts": 7429564843401.086,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843418.999, "dur": 4.732,
    "args": {
      "External id": 3799,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1386,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1386, "pid": 537, "tid": 537, "ts": 7429564843418.999,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843434.868, "dur": 6.764,
    "args": {
      "External id": 3802,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1387,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1387, "pid": 537, "tid": 537, "ts": 7429564843434.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843482.195, "dur": 0.185,
    "args": {
      "External id": 3814,      "cid": 370, "correlation": 1388
    }
  },
  {
    "ph": "f", "id": 1388, "pid": 537, "tid": 537, "ts": 7429564843482.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843485.408, "dur": 0.151,
    "args": {
      "External id": 3814,      "cid": 370, "correlation": 1389
    }
  },
  {
    "ph": "f", "id": 1389, "pid": 537, "tid": 537, "ts": 7429564843485.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843488.504, "dur": 0.169,
    "args": {
      "External id": 3814,      "cid": 370, "correlation": 1390
    }
  },
  {
    "ph": "f", "id": 1390, "pid": 537, "tid": 537, "ts": 7429564843488.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843490.420, "dur": 4.915,
    "args": {
      "External id": 3814,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1391,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1391, "pid": 537, "tid": 537, "ts": 7429564843490.420,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843511.311, "dur": 4.149,
    "args": {
      "External id": 3823,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1392,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1392, "pid": 537, "tid": 537, "ts": 7429564843511.311,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843558.451, "dur": 0.199,
    "args": {
      "External id": 3833,      "cid": 370, "correlation": 1393
    }
  },
  {
    "ph": "f", "id": 1393, "pid": 537, "tid": 537, "ts": 7429564843558.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843561.631, "dur": 0.154,
    "args": {
      "External id": 3833,      "cid": 370, "correlation": 1394
    }
  },
  {
    "ph": "f", "id": 1394, "pid": 537, "tid": 537, "ts": 7429564843561.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843563.629, "dur": 0.165,
    "args": {
      "External id": 3833,      "cid": 370, "correlation": 1395
    }
  },
  {
    "ph": "f", "id": 1395, "pid": 537, "tid": 537, "ts": 7429564843563.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843565.581, "dur": 4.449,
    "args": {
      "External id": 3833,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1396,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1396, "pid": 537, "tid": 537, "ts": 7429564843565.581,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564843581.987, "dur": 0.382,
    "args": {
      "External id": 3836,      "cid": 290, "correlation": 1397
    }
  },
  {
    "ph": "f", "id": 1397, "pid": 537, "tid": 537, "ts": 7429564843581.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564843582.821, "dur": 52.527,
    "args": {
      "External id": 3836,            "bytes": 2097152, "cid": 108, "correlation": 1398, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1398, "pid": 537, "tid": 537, "ts": 7429564843582.821,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843638.079, "dur": 5.243,
    "args": {
      "External id": 3836,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1399,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1399, "pid": 537, "tid": 537, "ts": 7429564843638.079,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843665.549, "dur": 5.767,
    "args": {
      "External id": 3838,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1400,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1400, "pid": 537, "tid": 537, "ts": 7429564843665.549,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843682.748, "dur": 6.551,
    "args": {
      "External id": 3841,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1401,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1401, "pid": 537, "tid": 537, "ts": 7429564843682.748,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843697.798, "dur": 4.105,
    "args": {
      "External id": 3842,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1402,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1402, "pid": 537, "tid": 537, "ts": 7429564843697.798,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843708.177, "dur": 4.939,
    "args": {
      "External id": 3843,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1403,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1403, "pid": 537, "tid": 537, "ts": 7429564843708.177,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843720.613, "dur": 4.972,
    "args": {
      "External id": 3844,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1404,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1404, "pid": 537, "tid": 537, "ts": 7429564843720.613,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843733.367, "dur": 5.517,
    "args": {
      "External id": 3846,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1405,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1405, "pid": 537, "tid": 537, "ts": 7429564843733.367,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843778.927, "dur": 0.306,
    "args": {
      "External id": 3854,      "cid": 370, "correlation": 1406
    }
  },
  {
    "ph": "f", "id": 1406, "pid": 537, "tid": 537, "ts": 7429564843778.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843782.874, "dur": 0.150,
    "args": {
      "External id": 3854,      "cid": 370, "correlation": 1407
    }
  },
  {
    "ph": "f", "id": 1407, "pid": 537, "tid": 537, "ts": 7429564843782.874,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843785.423, "dur": 0.157,
    "args": {
      "External id": 3854,      "cid": 370, "correlation": 1408
    }
  },
  {
    "ph": "f", "id": 1408, "pid": 537, "tid": 537, "ts": 7429564843785.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843793.162, "dur": 6.467,
    "args": {
      "External id": 3854,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1409,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1409, "pid": 537, "tid": 537, "ts": 7429564843793.162,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843800.132, "dur": 4.114,
    "args": {
      "External id": 3854,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1410,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1410, "pid": 537, "tid": 537, "ts": 7429564843800.132,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843821.212, "dur": 5.382,
    "args": {
      "External id": 3857,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1411,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1411, "pid": 537, "tid": 537, "ts": 7429564843821.212,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564843846.820, "dur": 0.331,
    "args": {
      "External id": 3865,      "cid": 290, "correlation": 1412
    }
  },
  {
    "ph": "f", "id": 1412, "pid": 537, "tid": 537, "ts": 7429564843846.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564843847.520, "dur": 36.545,
    "args": {
      "External id": 3865,            "bytes": 2097152, "cid": 108, "correlation": 1413, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1413, "pid": 537, "tid": 537, "ts": 7429564843847.520,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843896.431, "dur": 0.205,
    "args": {
      "External id": 3865,      "cid": 370, "correlation": 1414
    }
  },
  {
    "ph": "f", "id": 1414, "pid": 537, "tid": 537, "ts": 7429564843896.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843899.602, "dur": 0.149,
    "args": {
      "External id": 3865,      "cid": 370, "correlation": 1415
    }
  },
  {
    "ph": "f", "id": 1415, "pid": 537, "tid": 537, "ts": 7429564843899.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843901.887, "dur": 0.156,
    "args": {
      "External id": 3865,      "cid": 370, "correlation": 1416
    }
  },
  {
    "ph": "f", "id": 1416, "pid": 537, "tid": 537, "ts": 7429564843901.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843906.311, "dur": 6.784,
    "args": {
      "External id": 3865,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1417,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1417, "pid": 537, "tid": 537, "ts": 7429564843906.311,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843913.568, "dur": 3.637,
    "args": {
      "External id": 3865,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1418,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1418, "pid": 537, "tid": 537, "ts": 7429564843913.568,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843927.857, "dur": 4.344,
    "args": {
      "External id": 3868,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1419,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1419, "pid": 537, "tid": 537, "ts": 7429564843927.857,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843960.687, "dur": 0.159,
    "args": {
      "External id": 3876,      "cid": 370, "correlation": 1420
    }
  },
  {
    "ph": "f", "id": 1420, "pid": 537, "tid": 537, "ts": 7429564843960.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843963.952, "dur": 0.151,
    "args": {
      "External id": 3876,      "cid": 370, "correlation": 1421
    }
  },
  {
    "ph": "f", "id": 1421, "pid": 537, "tid": 537, "ts": 7429564843963.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564843965.957, "dur": 0.156,
    "args": {
      "External id": 3876,      "cid": 370, "correlation": 1422
    }
  },
  {
    "ph": "f", "id": 1422, "pid": 537, "tid": 537, "ts": 7429564843965.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843969.702, "dur": 5.866,
    "args": {
      "External id": 3876,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1423,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1423, "pid": 537, "tid": 537, "ts": 7429564843969.702,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843975.830, "dur": 4.285,
    "args": {
      "External id": 3876,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1424,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1424, "pid": 537, "tid": 537, "ts": 7429564843975.830,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564843989.956, "dur": 4.231,
    "args": {
      "External id": 3879,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1425,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1425, "pid": 537, "tid": 537, "ts": 7429564843989.956,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844021.482, "dur": 4.158,
    "args": {
      "External id": 3881,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1426,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1426, "pid": 537, "tid": 537, "ts": 7429564844021.482,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844034.261, "dur": 5.364,
    "args": {
      "External id": 3884,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1427,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1427, "pid": 537, "tid": 537, "ts": 7429564844034.261,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844047.964, "dur": 4.961,
    "args": {
      "External id": 3885,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1428,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1428, "pid": 537, "tid": 537, "ts": 7429564844047.964,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844057.761, "dur": 5.177,
    "args": {
      "External id": 3886,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1429,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1429, "pid": 537, "tid": 537, "ts": 7429564844057.761,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844069.915, "dur": 4.950,
    "args": {
      "External id": 3887,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1430,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1430, "pid": 537, "tid": 537, "ts": 7429564844069.915,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564844081.508, "dur": 0.209,
    "args": {
      "External id": 3889,      "cid": 290, "correlation": 1431
    }
  },
  {
    "ph": "f", "id": 1431, "pid": 537, "tid": 537, "ts": 7429564844081.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564844081.987, "dur": 43.280,
    "args": {
      "External id": 3889,            "bytes": 2097152, "cid": 108, "correlation": 1432, "ptr": "0x100000002"
    }
  },
  {
    "ph": "s", "id": 1432, "pid": 537, "tid": 537, "ts": 7429564844081.987,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844129.468, "dur": 4.435,
    "args": {
      "External id": 3889,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1433,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1433, "pid": 537, "tid": 537, "ts": 7429564844129.468,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844185.275, "dur": 0.258,
    "args": {
      "External id": 3897,      "cid": 370, "correlation": 1434
    }
  },
  {
    "ph": "f", "id": 1434, "pid": 537, "tid": 537, "ts": 7429564844185.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844189.457, "dur": 0.160,
    "args": {
      "External id": 3897,      "cid": 370, "correlation": 1435
    }
  },
  {
    "ph": "f", "id": 1435, "pid": 537, "tid": 537, "ts": 7429564844189.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844192.287, "dur": 0.152,
    "args": {
      "External id": 3897,      "cid": 370, "correlation": 1436
    }
  },
  {
    "ph": "f", "id": 1436, "pid": 537, "tid": 537, "ts": 7429564844192.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844195.326, "dur": 5.508,
    "args": {
      "External id": 3897,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1437,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1437, "pid": 537, "tid": 537, "ts": 7429564844195.326,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844239.745, "dur": 0.204,
    "args": {
      "External id": 3910,      "cid": 370, "correlation": 1438
    }
  },
  {
    "ph": "f", "id": 1438, "pid": 537, "tid": 537, "ts": 7429564844239.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844242.960, "dur": 0.168,
    "args": {
      "External id": 3910,      "cid": 370, "correlation": 1439
    }
  },
  {
    "ph": "f", "id": 1439, "pid": 537, "tid": 537, "ts": 7429564844242.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844245.123, "dur": 0.148,
    "args": {
      "External id": 3910,      "cid": 370, "correlation": 1440
    }
  },
  {
    "ph": "f", "id": 1440, "pid": 537, "tid": 537, "ts": 7429564844245.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844250.221, "dur": 6.644,
    "args": {
      "External id": 3910,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1441,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1441, "pid": 537, "tid": 537, "ts": 7429564844250.221,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844257.379, "dur": 3.886,
    "args": {
      "External id": 3910,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1442,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1442, "pid": 537, "tid": 537, "ts": 7429564844257.379,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844296.935, "dur": 0.210,
    "args": {
      "External id": 3923,      "cid": 370, "correlation": 1443
    }
  },
  {
    "ph": "f", "id": 1443, "pid": 537, "tid": 537, "ts": 7429564844296.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844299.720, "dur": 0.152,
    "args": {
      "External id": 3923,      "cid": 370, "correlation": 1444
    }
  },
  {
    "ph": "f", "id": 1444, "pid": 537, "tid": 537, "ts": 7429564844299.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844301.562, "dur": 0.145,
    "args": {
      "External id": 3923,      "cid": 370, "correlation": 1445
    }
  },
  {
    "ph": "f", "id": 1445, "pid": 537, "tid": 537, "ts": 7429564844301.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844305.280, "dur": 5.475,
    "args": {
      "External id": 3923,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1446,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1446, "pid": 537, "tid": 537, "ts": 7429564844305.280,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844311.040, "dur": 3.695,
    "args": {
      "External id": 3923,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1447,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1447, "pid": 537, "tid": 537, "ts": 7429564844311.040,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844339.127, "dur": 3.521,
    "args": {
      "External id": 3933,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1448,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1448, "pid": 537, "tid": 537, "ts": 7429564844339.127,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844367.836, "dur": 4.661,
    "args": {
      "External id": 3938,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1449,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1449, "pid": 537, "tid": 537, "ts": 7429564844367.836,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844386.809, "dur": 6.163,
    "args": {
      "External id": 3939,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1450,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1450, "pid": 537, "tid": 537, "ts": 7429564844386.809,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844401.332, "dur": 5.087,
    "args": {
      "External id": 3940,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1451,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1451, "pid": 537, "tid": 537, "ts": 7429564844401.332,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844412.199, "dur": 6.935,
    "args": {
      "External id": 3941,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1452,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1452, "pid": 537, "tid": 537, "ts": 7429564844412.199,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844425.194, "dur": 3.784,
    "args": {
      "External id": 3942,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1453,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1453, "pid": 537, "tid": 537, "ts": 7429564844425.194,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844442.338, "dur": 3.585,
    "args": {
      "External id": 3947,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1454,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1454, "pid": 537, "tid": 537, "ts": 7429564844442.338,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844453.365, "dur": 7.422,
    "args": {
      "External id": 3948,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1455,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1455, "pid": 537, "tid": 537, "ts": 7429564844453.365,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844466.803, "dur": 3.378,
    "args": {
      "External id": 3949,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1456,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1456, "pid": 537, "tid": 537, "ts": 7429564844466.803,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844475.193, "dur": 3.353,
    "args": {
      "External id": 3950,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1457,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1457, "pid": 537, "tid": 537, "ts": 7429564844475.193,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844539.130, "dur": 6.360,
    "args": {
      "External id": 3967,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1458,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1458, "pid": 537, "tid": 537, "ts": 7429564844539.130,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844576.989, "dur": 4.041,
    "args": {
      "External id": 3985,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1459,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1459, "pid": 537, "tid": 537, "ts": 7429564844576.989,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844615.598, "dur": 0.295,
    "args": {
      "External id": 3998,      "cid": 370, "correlation": 1460
    }
  },
  {
    "ph": "f", "id": 1460, "pid": 537, "tid": 537, "ts": 7429564844615.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844620.009, "dur": 0.164,
    "args": {
      "External id": 3998,      "cid": 370, "correlation": 1461
    }
  },
  {
    "ph": "f", "id": 1461, "pid": 537, "tid": 537, "ts": 7429564844620.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844622.416, "dur": 0.158,
    "args": {
      "External id": 3998,      "cid": 370, "correlation": 1462
    }
  },
  {
    "ph": "f", "id": 1462, "pid": 537, "tid": 537, "ts": 7429564844622.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844624.918, "dur": 4.566,
    "args": {
      "External id": 3998,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1463,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1463, "pid": 537, "tid": 537, "ts": 7429564844624.918,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844641.368, "dur": 5.204,
    "args": {
      "External id": 4001,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1464,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1464, "pid": 537, "tid": 537, "ts": 7429564844641.368,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844660.859, "dur": 3.761,
    "args": {
      "External id": 4008,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1465,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1465, "pid": 537, "tid": 537, "ts": 7429564844660.859,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844675.840, "dur": 5.349,
    "args": {
      "External id": 4011,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1466,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1466, "pid": 537, "tid": 537, "ts": 7429564844675.840,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844720.565, "dur": 0.194,
    "args": {
      "External id": 4023,      "cid": 370, "correlation": 1467
    }
  },
  {
    "ph": "f", "id": 1467, "pid": 537, "tid": 537, "ts": 7429564844720.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844724.404, "dur": 0.145,
    "args": {
      "External id": 4023,      "cid": 370, "correlation": 1468
    }
  },
  {
    "ph": "f", "id": 1468, "pid": 537, "tid": 537, "ts": 7429564844724.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844727.230, "dur": 0.161,
    "args": {
      "External id": 4023,      "cid": 370, "correlation": 1469
    }
  },
  {
    "ph": "f", "id": 1469, "pid": 537, "tid": 537, "ts": 7429564844727.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844729.143, "dur": 4.987,
    "args": {
      "External id": 4023,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1470,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1470, "pid": 537, "tid": 537, "ts": 7429564844729.143,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844750.066, "dur": 4.406,
    "args": {
      "External id": 4032,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1471,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1471, "pid": 537, "tid": 537, "ts": 7429564844750.066,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844793.234, "dur": 0.264,
    "args": {
      "External id": 4042,      "cid": 370, "correlation": 1472
    }
  },
  {
    "ph": "f", "id": 1472, "pid": 537, "tid": 537, "ts": 7429564844793.234,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844796.332, "dur": 0.153,
    "args": {
      "External id": 4042,      "cid": 370, "correlation": 1473
    }
  },
  {
    "ph": "f", "id": 1473, "pid": 537, "tid": 537, "ts": 7429564844796.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564844798.360, "dur": 0.194,
    "args": {
      "External id": 4042,      "cid": 370, "correlation": 1474
    }
  },
  {
    "ph": "f", "id": 1474, "pid": 537, "tid": 537, "ts": 7429564844798.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844800.208, "dur": 4.089,
    "args": {
      "External id": 4042,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1475,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1475, "pid": 537, "tid": 537, "ts": 7429564844800.208,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844816.992, "dur": 4.813,
    "args": {
      "External id": 4045,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1476,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1476, "pid": 537, "tid": 537, "ts": 7429564844816.992,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844839.690, "dur": 5.552,
    "args": {
      "External id": 4047,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1477,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1477, "pid": 537, "tid": 537, "ts": 7429564844839.690,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844855.739, "dur": 5.217,
    "args": {
      "External id": 4050,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1478,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1478, "pid": 537, "tid": 537, "ts": 7429564844855.739,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844869.127, "dur": 4.488,
    "args": {
      "External id": 4051,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1479,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1479, "pid": 537, "tid": 537, "ts": 7429564844869.127,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844880.508, "dur": 4.956,
    "args": {
      "External id": 4052,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1480,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1480, "pid": 537, "tid": 537, "ts": 7429564844880.508,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844891.845, "dur": 3.664,
    "args": {
      "External id": 4053,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1481,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1481, "pid": 537, "tid": 537, "ts": 7429564844891.845,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564844904.430, "dur": 4.756,
    "args": {
      "External id": 4055,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1482,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1482, "pid": 537, "tid": 537, "ts": 7429564844904.430,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564844933.729, "dur": 0.321,
    "args": {
      "External id": 4063,      "cid": 290, "correlation": 1483
    }
  },
  {
    "ph": "f", "id": 1483, "pid": 537, "tid": 537, "ts": 7429564844933.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564844934.620, "dur": 52.479,
    "args": {
      "External id": 4063,            "bytes": 2097152, "cid": 108, "correlation": 1484, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1484, "pid": 537, "tid": 537, "ts": 7429564844934.620,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845001.017, "dur": 0.227,
    "args": {
      "External id": 4063,      "cid": 370, "correlation": 1485
    }
  },
  {
    "ph": "f", "id": 1485, "pid": 537, "tid": 537, "ts": 7429564845001.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845005.040, "dur": 0.143,
    "args": {
      "External id": 4063,      "cid": 370, "correlation": 1486
    }
  },
  {
    "ph": "f", "id": 1486, "pid": 537, "tid": 537, "ts": 7429564845005.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845007.520, "dur": 0.163,
    "args": {
      "External id": 4063,      "cid": 370, "correlation": 1487
    }
  },
  {
    "ph": "f", "id": 1487, "pid": 537, "tid": 537, "ts": 7429564845007.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845014.826, "dur": 4.991,
    "args": {
      "External id": 4063,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1488,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1488, "pid": 537, "tid": 537, "ts": 7429564845014.826,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845020.200, "dur": 4.765,
    "args": {
      "External id": 4063,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1489,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1489, "pid": 537, "tid": 537, "ts": 7429564845020.200,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845044.232, "dur": 5.023,
    "args": {
      "External id": 4066,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1490,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1490, "pid": 537, "tid": 537, "ts": 7429564845044.232,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564845069.539, "dur": 0.267,
    "args": {
      "External id": 4074,      "cid": 290, "correlation": 1491
    }
  },
  {
    "ph": "f", "id": 1491, "pid": 537, "tid": 537, "ts": 7429564845069.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564845070.247, "dur": 29.997,
    "args": {
      "External id": 4074,            "bytes": 2097152, "cid": 108, "correlation": 1492, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1492, "pid": 537, "tid": 537, "ts": 7429564845070.247,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845111.554, "dur": 0.224,
    "args": {
      "External id": 4074,      "cid": 370, "correlation": 1493
    }
  },
  {
    "ph": "f", "id": 1493, "pid": 537, "tid": 537, "ts": 7429564845111.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845115.169, "dur": 0.157,
    "args": {
      "External id": 4074,      "cid": 370, "correlation": 1494
    }
  },
  {
    "ph": "f", "id": 1494, "pid": 537, "tid": 537, "ts": 7429564845115.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845117.355, "dur": 0.157,
    "args": {
      "External id": 4074,      "cid": 370, "correlation": 1495
    }
  },
  {
    "ph": "f", "id": 1495, "pid": 537, "tid": 537, "ts": 7429564845117.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845121.981, "dur": 6.220,
    "args": {
      "External id": 4074,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1496,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1496, "pid": 537, "tid": 537, "ts": 7429564845121.981,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845128.539, "dur": 3.282,
    "args": {
      "External id": 4074,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1497,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1497, "pid": 537, "tid": 537, "ts": 7429564845128.539,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845143.415, "dur": 4.170,
    "args": {
      "External id": 4077,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1498,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1498, "pid": 537, "tid": 537, "ts": 7429564845143.415,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845174.825, "dur": 0.193,
    "args": {
      "External id": 4085,      "cid": 370, "correlation": 1499
    }
  },
  {
    "ph": "f", "id": 1499, "pid": 537, "tid": 537, "ts": 7429564845174.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845178.112, "dur": 0.144,
    "args": {
      "External id": 4085,      "cid": 370, "correlation": 1500
    }
  },
  {
    "ph": "f", "id": 1500, "pid": 537, "tid": 537, "ts": 7429564845178.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845180.043, "dur": 0.149,
    "args": {
      "External id": 4085,      "cid": 370, "correlation": 1501
    }
  },
  {
    "ph": "f", "id": 1501, "pid": 537, "tid": 537, "ts": 7429564845180.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845183.957, "dur": 4.181,
    "args": {
      "External id": 4085,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1502,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1502, "pid": 537, "tid": 537, "ts": 7429564845183.957,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845188.434, "dur": 3.352,
    "args": {
      "External id": 4085,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1503,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1503, "pid": 537, "tid": 537, "ts": 7429564845188.434,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845201.502, "dur": 5.901,
    "args": {
      "External id": 4088,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1504,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1504, "pid": 537, "tid": 537, "ts": 7429564845201.502,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845237.077, "dur": 5.004,
    "args": {
      "External id": 4090,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1505,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1505, "pid": 537, "tid": 537, "ts": 7429564845237.077,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845252.042, "dur": 4.781,
    "args": {
      "External id": 4093,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1506,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1506, "pid": 537, "tid": 537, "ts": 7429564845252.042,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845266.361, "dur": 3.469,
    "args": {
      "External id": 4094,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1507,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1507, "pid": 537, "tid": 537, "ts": 7429564845266.361,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845275.184, "dur": 4.545,
    "args": {
      "External id": 4095,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1508,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1508, "pid": 537, "tid": 537, "ts": 7429564845275.184,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845286.418, "dur": 5.470,
    "args": {
      "External id": 4096,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1509,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1509, "pid": 537, "tid": 537, "ts": 7429564845286.418,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845332.214, "dur": 4.277,
    "args": {
      "External id": 4098,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1510,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1510, "pid": 537, "tid": 537, "ts": 7429564845332.214,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845375.519, "dur": 0.325,
    "args": {
      "External id": 4106,      "cid": 370, "correlation": 1511
    }
  },
  {
    "ph": "f", "id": 1511, "pid": 537, "tid": 537, "ts": 7429564845375.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845378.857, "dur": 0.155,
    "args": {
      "External id": 4106,      "cid": 370, "correlation": 1512
    }
  },
  {
    "ph": "f", "id": 1512, "pid": 537, "tid": 537, "ts": 7429564845378.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845381.139, "dur": 0.166,
    "args": {
      "External id": 4106,      "cid": 370, "correlation": 1513
    }
  },
  {
    "ph": "f", "id": 1513, "pid": 537, "tid": 537, "ts": 7429564845381.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845383.397, "dur": 6.400,
    "args": {
      "External id": 4106,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1514,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1514, "pid": 537, "tid": 537, "ts": 7429564845383.397,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845426.317, "dur": 0.188,
    "args": {
      "External id": 4119,      "cid": 370, "correlation": 1515
    }
  },
  {
    "ph": "f", "id": 1515, "pid": 537, "tid": 537, "ts": 7429564845426.317,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845429.866, "dur": 0.154,
    "args": {
      "External id": 4119,      "cid": 370, "correlation": 1516
    }
  },
  {
    "ph": "f", "id": 1516, "pid": 537, "tid": 537, "ts": 7429564845429.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845431.785, "dur": 0.174,
    "args": {
      "External id": 4119,      "cid": 370, "correlation": 1517
    }
  },
  {
    "ph": "f", "id": 1517, "pid": 537, "tid": 537, "ts": 7429564845431.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845436.107, "dur": 5.338,
    "args": {
      "External id": 4119,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1518,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1518, "pid": 537, "tid": 537, "ts": 7429564845436.107,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845441.837, "dur": 4.186,
    "args": {
      "External id": 4119,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1519,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1519, "pid": 537, "tid": 537, "ts": 7429564845441.837,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845501.011, "dur": 0.193,
    "args": {
      "External id": 4132,      "cid": 370, "correlation": 1520
    }
  },
  {
    "ph": "f", "id": 1520, "pid": 537, "tid": 537, "ts": 7429564845501.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845503.324, "dur": 0.145,
    "args": {
      "External id": 4132,      "cid": 370, "correlation": 1521
    }
  },
  {
    "ph": "f", "id": 1521, "pid": 537, "tid": 537, "ts": 7429564845503.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845505.173, "dur": 0.142,
    "args": {
      "External id": 4132,      "cid": 370, "correlation": 1522
    }
  },
  {
    "ph": "f", "id": 1522, "pid": 537, "tid": 537, "ts": 7429564845505.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845508.618, "dur": 4.365,
    "args": {
      "External id": 4132,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1523,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1523, "pid": 537, "tid": 537, "ts": 7429564845508.618,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845513.426, "dur": 3.524,
    "args": {
      "External id": 4132,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1524,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1524, "pid": 537, "tid": 537, "ts": 7429564845513.426,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845540.670, "dur": 4.161,
    "args": {
      "External id": 4142,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1525,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1525, "pid": 537, "tid": 537, "ts": 7429564845540.670,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845566.763, "dur": 4.801,
    "args": {
      "External id": 4147,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1526,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1526, "pid": 537, "tid": 537, "ts": 7429564845566.763,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564845582.585, "dur": 0.292,
    "args": {
      "External id": 4148,      "cid": 290, "correlation": 1527
    }
  },
  {
    "ph": "f", "id": 1527, "pid": 537, "tid": 537, "ts": 7429564845582.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564845583.261, "dur": 41.127,
    "args": {
      "External id": 4148,            "bytes": 2097152, "cid": 108, "correlation": 1528, "ptr": "0x5569ba9ecb70"
    }
  },
  {
    "ph": "s", "id": 1528, "pid": 537, "tid": 537, "ts": 7429564845583.261,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845628.756, "dur": 6.790,
    "args": {
      "External id": 4148,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1529,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1529, "pid": 537, "tid": 537, "ts": 7429564845628.756,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845644.306, "dur": 3.940,
    "args": {
      "External id": 4149,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1530,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1530, "pid": 537, "tid": 537, "ts": 7429564845644.306,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845654.504, "dur": 4.330,
    "args": {
      "External id": 4150,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1531,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1531, "pid": 537, "tid": 537, "ts": 7429564845654.504,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845665.156, "dur": 3.267,
    "args": {
      "External id": 4151,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1532,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1532, "pid": 537, "tid": 537, "ts": 7429564845665.156,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845682.133, "dur": 4.283,
    "args": {
      "External id": 4156,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1533,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1533, "pid": 537, "tid": 537, "ts": 7429564845682.133,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845693.681, "dur": 4.988,
    "args": {
      "External id": 4157,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1534,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1534, "pid": 537, "tid": 537, "ts": 7429564845693.681,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845704.067, "dur": 3.454,
    "args": {
      "External id": 4158,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1535,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1535, "pid": 537, "tid": 537, "ts": 7429564845704.067,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845712.345, "dur": 3.078,
    "args": {
      "External id": 4159,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1536,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1536, "pid": 537, "tid": 537, "ts": 7429564845712.345,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845775.586, "dur": 4.818,
    "args": {
      "External id": 4176,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1537,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1537, "pid": 537, "tid": 537, "ts": 7429564845775.586,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845808.243, "dur": 3.546,
    "args": {
      "External id": 4194,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1538,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1538, "pid": 537, "tid": 537, "ts": 7429564845808.243,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845848.545, "dur": 0.312,
    "args": {
      "External id": 4207,      "cid": 370, "correlation": 1539
    }
  },
  {
    "ph": "f", "id": 1539, "pid": 537, "tid": 537, "ts": 7429564845848.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845853.722, "dur": 0.161,
    "args": {
      "External id": 4207,      "cid": 370, "correlation": 1540
    }
  },
  {
    "ph": "f", "id": 1540, "pid": 537, "tid": 537, "ts": 7429564845853.722,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845856.430, "dur": 0.186,
    "args": {
      "External id": 4207,      "cid": 370, "correlation": 1541
    }
  },
  {
    "ph": "f", "id": 1541, "pid": 537, "tid": 537, "ts": 7429564845856.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845859.491, "dur": 5.342,
    "args": {
      "External id": 4207,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1542,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1542, "pid": 537, "tid": 537, "ts": 7429564845859.491,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845877.307, "dur": 4.921,
    "args": {
      "External id": 4210,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1543,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1543, "pid": 537, "tid": 537, "ts": 7429564845877.307,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845896.752, "dur": 3.980,
    "args": {
      "External id": 4217,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1544,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1544, "pid": 537, "tid": 537, "ts": 7429564845896.752,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845912.503, "dur": 6.478,
    "args": {
      "External id": 4220,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1545,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1545, "pid": 537, "tid": 537, "ts": 7429564845912.503,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845961.461, "dur": 0.185,
    "args": {
      "External id": 4232,      "cid": 370, "correlation": 1546
    }
  },
  {
    "ph": "f", "id": 1546, "pid": 537, "tid": 537, "ts": 7429564845961.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845965.026, "dur": 0.163,
    "args": {
      "External id": 4232,      "cid": 370, "correlation": 1547
    }
  },
  {
    "ph": "f", "id": 1547, "pid": 537, "tid": 537, "ts": 7429564845965.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564845967.973, "dur": 0.151,
    "args": {
      "External id": 4232,      "cid": 370, "correlation": 1548
    }
  },
  {
    "ph": "f", "id": 1548, "pid": 537, "tid": 537, "ts": 7429564845967.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845970.156, "dur": 5.209,
    "args": {
      "External id": 4232,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1549,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1549, "pid": 537, "tid": 537, "ts": 7429564845970.156,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564845992.042, "dur": 4.008,
    "args": {
      "External id": 4241,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1550,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1550, "pid": 537, "tid": 537, "ts": 7429564845992.042,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846036.710, "dur": 0.284,
    "args": {
      "External id": 4251,      "cid": 370, "correlation": 1551
    }
  },
  {
    "ph": "f", "id": 1551, "pid": 537, "tid": 537, "ts": 7429564846036.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846039.740, "dur": 0.168,
    "args": {
      "External id": 4251,      "cid": 370, "correlation": 1552
    }
  },
  {
    "ph": "f", "id": 1552, "pid": 537, "tid": 537, "ts": 7429564846039.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846041.898, "dur": 0.200,
    "args": {
      "External id": 4251,      "cid": 370, "correlation": 1553
    }
  },
  {
    "ph": "f", "id": 1553, "pid": 537, "tid": 537, "ts": 7429564846041.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846043.805, "dur": 4.695,
    "args": {
      "External id": 4251,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1554,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1554, "pid": 537, "tid": 537, "ts": 7429564846043.805,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846060.587, "dur": 5.203,
    "args": {
      "External id": 4254,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1555,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1555, "pid": 537, "tid": 537, "ts": 7429564846060.587,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846108.160, "dur": 5.516,
    "args": {
      "External id": 4256,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1556,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1556, "pid": 537, "tid": 537, "ts": 7429564846108.160,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846125.820, "dur": 5.599,
    "args": {
      "External id": 4259,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1557,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1557, "pid": 537, "tid": 537, "ts": 7429564846125.820,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846142.051, "dur": 4.383,
    "args": {
      "External id": 4260,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1558,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1558, "pid": 537, "tid": 537, "ts": 7429564846142.051,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846153.328, "dur": 4.609,
    "args": {
      "External id": 4261,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1559,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1559, "pid": 537, "tid": 537, "ts": 7429564846153.328,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846166.042, "dur": 4.286,
    "args": {
      "External id": 4262,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1560,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1560, "pid": 537, "tid": 537, "ts": 7429564846166.042,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846178.025, "dur": 6.056,
    "args": {
      "External id": 4264,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1561,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1561, "pid": 537, "tid": 537, "ts": 7429564846178.025,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846224.711, "dur": 0.209,
    "args": {
      "External id": 4272,      "cid": 370, "correlation": 1562
    }
  },
  {
    "ph": "f", "id": 1562, "pid": 537, "tid": 537, "ts": 7429564846224.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846228.673, "dur": 0.145,
    "args": {
      "External id": 4272,      "cid": 370, "correlation": 1563
    }
  },
  {
    "ph": "f", "id": 1563, "pid": 537, "tid": 537, "ts": 7429564846228.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846231.414, "dur": 0.166,
    "args": {
      "External id": 4272,      "cid": 370, "correlation": 1564
    }
  },
  {
    "ph": "f", "id": 1564, "pid": 537, "tid": 537, "ts": 7429564846231.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846239.134, "dur": 5.131,
    "args": {
      "External id": 4272,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1565,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1565, "pid": 537, "tid": 537, "ts": 7429564846239.134,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846244.813, "dur": 5.322,
    "args": {
      "External id": 4272,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1566,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1566, "pid": 537, "tid": 537, "ts": 7429564846244.813,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564846267.140, "dur": 0.355,
    "args": {
      "External id": 4275,      "cid": 290, "correlation": 1567
    }
  },
  {
    "ph": "f", "id": 1567, "pid": 537, "tid": 537, "ts": 7429564846267.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564846267.897, "dur": 54.636,
    "args": {
      "External id": 4275,            "bytes": 2097152, "cid": 108, "correlation": 1568, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 1568, "pid": 537, "tid": 537, "ts": 7429564846267.897,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846325.783, "dur": 5.257,
    "args": {
      "External id": 4275,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1569,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1569, "pid": 537, "tid": 537, "ts": 7429564846325.783,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846362.557, "dur": 0.189,
    "args": {
      "External id": 4283,      "cid": 370, "correlation": 1570
    }
  },
  {
    "ph": "f", "id": 1570, "pid": 537, "tid": 537, "ts": 7429564846362.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846366.315, "dur": 0.141,
    "args": {
      "External id": 4283,      "cid": 370, "correlation": 1571
    }
  },
  {
    "ph": "f", "id": 1571, "pid": 537, "tid": 537, "ts": 7429564846366.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846368.830, "dur": 0.145,
    "args": {
      "External id": 4283,      "cid": 370, "correlation": 1572
    }
  },
  {
    "ph": "f", "id": 1572, "pid": 537, "tid": 537, "ts": 7429564846368.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846373.752, "dur": 5.865,
    "args": {
      "External id": 4283,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1573,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1573, "pid": 537, "tid": 537, "ts": 7429564846373.752,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846380.114, "dur": 4.156,
    "args": {
      "External id": 4283,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1574,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1574, "pid": 537, "tid": 537, "ts": 7429564846380.114,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564846392.714, "dur": 0.294,
    "args": {
      "External id": 4286,      "cid": 290, "correlation": 1575
    }
  },
  {
    "ph": "f", "id": 1575, "pid": 537, "tid": 537, "ts": 7429564846392.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564846393.326, "dur": 27.618,
    "args": {
      "External id": 4286,            "bytes": 2097152, "cid": 108, "correlation": 1576, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 1576, "pid": 537, "tid": 537, "ts": 7429564846393.326,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846423.302, "dur": 4.871,
    "args": {
      "External id": 4286,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1577,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1577, "pid": 537, "tid": 537, "ts": 7429564846423.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846456.490, "dur": 0.197,
    "args": {
      "External id": 4294,      "cid": 370, "correlation": 1578
    }
  },
  {
    "ph": "f", "id": 1578, "pid": 537, "tid": 537, "ts": 7429564846456.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846459.697, "dur": 0.156,
    "args": {
      "External id": 4294,      "cid": 370, "correlation": 1579
    }
  },
  {
    "ph": "f", "id": 1579, "pid": 537, "tid": 537, "ts": 7429564846459.697,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846461.995, "dur": 0.144,
    "args": {
      "External id": 4294,      "cid": 370, "correlation": 1580
    }
  },
  {
    "ph": "f", "id": 1580, "pid": 537, "tid": 537, "ts": 7429564846461.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846465.917, "dur": 4.989,
    "args": {
      "External id": 4294,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1581,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1581, "pid": 537, "tid": 537, "ts": 7429564846465.917,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846471.400, "dur": 5.442,
    "args": {
      "External id": 4294,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1582,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1582, "pid": 537, "tid": 537, "ts": 7429564846471.400,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846487.001, "dur": 4.781,
    "args": {
      "External id": 4297,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1583,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1583, "pid": 537, "tid": 537, "ts": 7429564846487.001,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846521.091, "dur": 4.379,
    "args": {
      "External id": 4299,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1584,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1584, "pid": 537, "tid": 537, "ts": 7429564846521.091,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846535.141, "dur": 6.190,
    "args": {
      "External id": 4302,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1585,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1585, "pid": 537, "tid": 537, "ts": 7429564846535.141,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846549.966, "dur": 5.318,
    "args": {
      "External id": 4303,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1586,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1586, "pid": 537, "tid": 537, "ts": 7429564846549.966,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846563.224, "dur": 4.571,
    "args": {
      "External id": 4304,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1587,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1587, "pid": 537, "tid": 537, "ts": 7429564846563.224,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846574.933, "dur": 4.174,
    "args": {
      "External id": 4305,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1588,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1588, "pid": 537, "tid": 537, "ts": 7429564846574.933,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846587.133, "dur": 3.608,
    "args": {
      "External id": 4307,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1589,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1589, "pid": 537, "tid": 537, "ts": 7429564846587.133,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846633.136, "dur": 0.277,
    "args": {
      "External id": 4315,      "cid": 370, "correlation": 1590
    }
  },
  {
    "ph": "f", "id": 1590, "pid": 537, "tid": 537, "ts": 7429564846633.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846637.103, "dur": 0.177,
    "args": {
      "External id": 4315,      "cid": 370, "correlation": 1591
    }
  },
  {
    "ph": "f", "id": 1591, "pid": 537, "tid": 537, "ts": 7429564846637.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846639.665, "dur": 0.174,
    "args": {
      "External id": 4315,      "cid": 370, "correlation": 1592
    }
  },
  {
    "ph": "f", "id": 1592, "pid": 537, "tid": 537, "ts": 7429564846639.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846642.082, "dur": 5.066,
    "args": {
      "External id": 4315,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1593,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1593, "pid": 537, "tid": 537, "ts": 7429564846642.082,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846688.120, "dur": 0.268,
    "args": {
      "External id": 4328,      "cid": 370, "correlation": 1594
    }
  },
  {
    "ph": "f", "id": 1594, "pid": 537, "tid": 537, "ts": 7429564846688.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846691.245, "dur": 0.153,
    "args": {
      "External id": 4328,      "cid": 370, "correlation": 1595
    }
  },
  {
    "ph": "f", "id": 1595, "pid": 537, "tid": 537, "ts": 7429564846691.245,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846693.512, "dur": 0.164,
    "args": {
      "External id": 4328,      "cid": 370, "correlation": 1596
    }
  },
  {
    "ph": "f", "id": 1596, "pid": 537, "tid": 537, "ts": 7429564846693.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846698.195, "dur": 6.105,
    "args": {
      "External id": 4328,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1597,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1597, "pid": 537, "tid": 537, "ts": 7429564846698.195,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846704.883, "dur": 4.585,
    "args": {
      "External id": 4328,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1598,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1598, "pid": 537, "tid": 537, "ts": 7429564846704.883,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846745.166, "dur": 0.195,
    "args": {
      "External id": 4341,      "cid": 370, "correlation": 1599
    }
  },
  {
    "ph": "f", "id": 1599, "pid": 537, "tid": 537, "ts": 7429564846745.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846747.599, "dur": 0.146,
    "args": {
      "External id": 4341,      "cid": 370, "correlation": 1600
    }
  },
  {
    "ph": "f", "id": 1600, "pid": 537, "tid": 537, "ts": 7429564846747.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564846749.386, "dur": 0.171,
    "args": {
      "External id": 4341,      "cid": 370, "correlation": 1601
    }
  },
  {
    "ph": "f", "id": 1601, "pid": 537, "tid": 537, "ts": 7429564846749.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846752.902, "dur": 4.534,
    "args": {
      "External id": 4341,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1602,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1602, "pid": 537, "tid": 537, "ts": 7429564846752.902,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846757.717, "dur": 3.335,
    "args": {
      "External id": 4341,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1603,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1603, "pid": 537, "tid": 537, "ts": 7429564846757.717,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846784.341, "dur": 3.643,
    "args": {
      "External id": 4351,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1604,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1604, "pid": 537, "tid": 537, "ts": 7429564846784.341,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846810.834, "dur": 5.692,
    "args": {
      "External id": 4356,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1605,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1605, "pid": 537, "tid": 537, "ts": 7429564846810.834,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846831.483, "dur": 6.209,
    "args": {
      "External id": 4357,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1606,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1606, "pid": 537, "tid": 537, "ts": 7429564846831.483,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846845.534, "dur": 4.899,
    "args": {
      "External id": 4358,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1607,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1607, "pid": 537, "tid": 537, "ts": 7429564846845.534,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846856.243, "dur": 5.058,
    "args": {
      "External id": 4359,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1608,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1608, "pid": 537, "tid": 537, "ts": 7429564846856.243,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846867.389, "dur": 3.313,
    "args": {
      "External id": 4360,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1609,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1609, "pid": 537, "tid": 537, "ts": 7429564846867.389,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846884.280, "dur": 4.275,
    "args": {
      "External id": 4365,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1610,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1610, "pid": 537, "tid": 537, "ts": 7429564846884.280,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846894.377, "dur": 5.015,
    "args": {
      "External id": 4366,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1611,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1611, "pid": 537, "tid": 537, "ts": 7429564846894.377,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846905.605, "dur": 5.648,
    "args": {
      "External id": 4367,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1612,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1612, "pid": 537, "tid": 537, "ts": 7429564846905.605,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846916.529, "dur": 4.543,
    "args": {
      "External id": 4368,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1613,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1613, "pid": 537, "tid": 537, "ts": 7429564846916.529,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564846982.709, "dur": 4.965,
    "args": {
      "External id": 4385,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1614,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1614, "pid": 537, "tid": 537, "ts": 7429564846982.709,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847017.694, "dur": 4.418,
    "args": {
      "External id": 4403,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1615,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1615, "pid": 537, "tid": 537, "ts": 7429564847017.694,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847059.226, "dur": 0.321,
    "args": {
      "External id": 4416,      "cid": 370, "correlation": 1616
    }
  },
  {
    "ph": "f", "id": 1616, "pid": 537, "tid": 537, "ts": 7429564847059.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847063.485, "dur": 0.146,
    "args": {
      "External id": 4416,      "cid": 370, "correlation": 1617
    }
  },
  {
    "ph": "f", "id": 1617, "pid": 537, "tid": 537, "ts": 7429564847063.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847066.167, "dur": 0.154,
    "args": {
      "External id": 4416,      "cid": 370, "correlation": 1618
    }
  },
  {
    "ph": "f", "id": 1618, "pid": 537, "tid": 537, "ts": 7429564847066.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847068.679, "dur": 4.856,
    "args": {
      "External id": 4416,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1619,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1619, "pid": 537, "tid": 537, "ts": 7429564847068.679,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847085.496, "dur": 4.351,
    "args": {
      "External id": 4419,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1620,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1620, "pid": 537, "tid": 537, "ts": 7429564847085.496,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847103.861, "dur": 4.992,
    "args": {
      "External id": 4426,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1621,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1621, "pid": 537, "tid": 537, "ts": 7429564847103.861,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847120.542, "dur": 6.601,
    "args": {
      "External id": 4429,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1622,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1622, "pid": 537, "tid": 537, "ts": 7429564847120.542,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847168.598, "dur": 0.203,
    "args": {
      "External id": 4441,      "cid": 370, "correlation": 1623
    }
  },
  {
    "ph": "f", "id": 1623, "pid": 537, "tid": 537, "ts": 7429564847168.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847171.998, "dur": 0.140,
    "args": {
      "External id": 4441,      "cid": 370, "correlation": 1624
    }
  },
  {
    "ph": "f", "id": 1624, "pid": 537, "tid": 537, "ts": 7429564847171.998,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847174.463, "dur": 0.150,
    "args": {
      "External id": 4441,      "cid": 370, "correlation": 1625
    }
  },
  {
    "ph": "f", "id": 1625, "pid": 537, "tid": 537, "ts": 7429564847174.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847176.547, "dur": 6.159,
    "args": {
      "External id": 4441,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1626,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1626, "pid": 537, "tid": 537, "ts": 7429564847176.547,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847199.943, "dur": 3.911,
    "args": {
      "External id": 4450,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1627,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1627, "pid": 537, "tid": 537, "ts": 7429564847199.943,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847246.244, "dur": 0.173,
    "args": {
      "External id": 4460,      "cid": 370, "correlation": 1628
    }
  },
  {
    "ph": "f", "id": 1628, "pid": 537, "tid": 537, "ts": 7429564847246.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847249.226, "dur": 0.159,
    "args": {
      "External id": 4460,      "cid": 370, "correlation": 1629
    }
  },
  {
    "ph": "f", "id": 1629, "pid": 537, "tid": 537, "ts": 7429564847249.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847251.495, "dur": 0.202,
    "args": {
      "External id": 4460,      "cid": 370, "correlation": 1630
    }
  },
  {
    "ph": "f", "id": 1630, "pid": 537, "tid": 537, "ts": 7429564847251.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847253.328, "dur": 4.707,
    "args": {
      "External id": 4460,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1631,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1631, "pid": 537, "tid": 537, "ts": 7429564847253.328,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847270.219, "dur": 5.150,
    "args": {
      "External id": 4463,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1632,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1632, "pid": 537, "tid": 537, "ts": 7429564847270.219,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847293.249, "dur": 5.164,
    "args": {
      "External id": 4465,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1633,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1633, "pid": 537, "tid": 537, "ts": 7429564847293.249,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847308.339, "dur": 5.399,
    "args": {
      "External id": 4468,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1634,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1634, "pid": 537, "tid": 537, "ts": 7429564847308.339,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847322.053, "dur": 4.238,
    "args": {
      "External id": 4469,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1635,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1635, "pid": 537, "tid": 537, "ts": 7429564847322.053,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847331.961, "dur": 4.158,
    "args": {
      "External id": 4470,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1636,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1636, "pid": 537, "tid": 537, "ts": 7429564847331.961,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847342.065, "dur": 4.053,
    "args": {
      "External id": 4471,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1637,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1637, "pid": 537, "tid": 537, "ts": 7429564847342.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564847354.973, "dur": 0.361,
    "args": {
      "External id": 4473,      "cid": 290, "correlation": 1638
    }
  },
  {
    "ph": "f", "id": 1638, "pid": 537, "tid": 537, "ts": 7429564847354.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564847355.741, "dur": 54.594,
    "args": {
      "External id": 4473,            "bytes": 2097152, "cid": 108, "correlation": 1639, "ptr": "0x100000002"
    }
  },
  {
    "ph": "s", "id": 1639, "pid": 537, "tid": 537, "ts": 7429564847355.741,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847412.464, "dur": 5.935,
    "args": {
      "External id": 4473,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1640,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1640, "pid": 537, "tid": 537, "ts": 7429564847412.464,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847462.635, "dur": 0.237,
    "args": {
      "External id": 4481,      "cid": 370, "correlation": 1641
    }
  },
  {
    "ph": "f", "id": 1641, "pid": 537, "tid": 537, "ts": 7429564847462.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847466.745, "dur": 0.148,
    "args": {
      "External id": 4481,      "cid": 370, "correlation": 1642
    }
  },
  {
    "ph": "f", "id": 1642, "pid": 537, "tid": 537, "ts": 7429564847466.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847469.224, "dur": 0.157,
    "args": {
      "External id": 4481,      "cid": 370, "correlation": 1643
    }
  },
  {
    "ph": "f", "id": 1643, "pid": 537, "tid": 537, "ts": 7429564847469.224,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847477.453, "dur": 5.594,
    "args": {
      "External id": 4481,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1644,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1644, "pid": 537, "tid": 537, "ts": 7429564847477.453,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847483.481, "dur": 4.105,
    "args": {
      "External id": 4481,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1645,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1645, "pid": 537, "tid": 537, "ts": 7429564847483.481,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847505.277, "dur": 4.962,
    "args": {
      "External id": 4484,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1646,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1646, "pid": 537, "tid": 537, "ts": 7429564847505.277,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564847528.547, "dur": 0.270,
    "args": {
      "External id": 4492,      "cid": 290, "correlation": 1647
    }
  },
  {
    "ph": "f", "id": 1647, "pid": 537, "tid": 537, "ts": 7429564847528.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564847529.155, "dur": 32.671,
    "args": {
      "External id": 4492,            "bytes": 2097152, "cid": 108, "correlation": 1648, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1648, "pid": 537, "tid": 537, "ts": 7429564847529.155,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847573.167, "dur": 0.190,
    "args": {
      "External id": 4492,      "cid": 370, "correlation": 1649
    }
  },
  {
    "ph": "f", "id": 1649, "pid": 537, "tid": 537, "ts": 7429564847573.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847576.842, "dur": 0.157,
    "args": {
      "External id": 4492,      "cid": 370, "correlation": 1650
    }
  },
  {
    "ph": "f", "id": 1650, "pid": 537, "tid": 537, "ts": 7429564847576.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847579.015, "dur": 0.171,
    "args": {
      "External id": 4492,      "cid": 370, "correlation": 1651
    }
  },
  {
    "ph": "f", "id": 1651, "pid": 537, "tid": 537, "ts": 7429564847579.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847583.773, "dur": 6.400,
    "args": {
      "External id": 4492,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1652,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1652, "pid": 537, "tid": 537, "ts": 7429564847583.773,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847590.495, "dur": 4.158,
    "args": {
      "External id": 4492,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1653,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1653, "pid": 537, "tid": 537, "ts": 7429564847590.495,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847605.338, "dur": 4.060,
    "args": {
      "External id": 4495,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1654,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1654, "pid": 537, "tid": 537, "ts": 7429564847605.338,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847636.681, "dur": 0.146,
    "args": {
      "External id": 4503,      "cid": 370, "correlation": 1655
    }
  },
  {
    "ph": "f", "id": 1655, "pid": 537, "tid": 537, "ts": 7429564847636.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847639.700, "dur": 0.147,
    "args": {
      "External id": 4503,      "cid": 370, "correlation": 1656
    }
  },
  {
    "ph": "f", "id": 1656, "pid": 537, "tid": 537, "ts": 7429564847639.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847641.501, "dur": 0.149,
    "args": {
      "External id": 4503,      "cid": 370, "correlation": 1657
    }
  },
  {
    "ph": "f", "id": 1657, "pid": 537, "tid": 537, "ts": 7429564847641.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847645.497, "dur": 4.706,
    "args": {
      "External id": 4503,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1658,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1658, "pid": 537, "tid": 537, "ts": 7429564847645.497,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847650.501, "dur": 4.505,
    "args": {
      "External id": 4503,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1659,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1659, "pid": 537, "tid": 537, "ts": 7429564847650.501,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847666.002, "dur": 3.889,
    "args": {
      "External id": 4506,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1660,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1660, "pid": 537, "tid": 537, "ts": 7429564847666.002,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847698.029, "dur": 4.028,
    "args": {
      "External id": 4508,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1661,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1661, "pid": 537, "tid": 537, "ts": 7429564847698.029,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847712.389, "dur": 5.528,
    "args": {
      "External id": 4511,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1662,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1662, "pid": 537, "tid": 537, "ts": 7429564847712.389,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847725.638, "dur": 3.794,
    "args": {
      "External id": 4512,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1663,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1663, "pid": 537, "tid": 537, "ts": 7429564847725.638,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847734.816, "dur": 5.667,
    "args": {
      "External id": 4513,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1664,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1664, "pid": 537, "tid": 537, "ts": 7429564847734.816,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847746.538, "dur": 4.784,
    "args": {
      "External id": 4514,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1665,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1665, "pid": 537, "tid": 537, "ts": 7429564847746.538,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847759.618, "dur": 4.310,
    "args": {
      "External id": 4516,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1666,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1666, "pid": 537, "tid": 537, "ts": 7429564847759.618,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564847793.567, "dur": 0.327,
    "args": {
      "External id": 4524,      "cid": 290, "correlation": 1667
    }
  },
  {
    "ph": "f", "id": 1667, "pid": 537, "tid": 537, "ts": 7429564847793.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564847794.461, "dur": 32.378,
    "args": {
      "External id": 4524,            "bytes": 2097152, "cid": 108, "correlation": 1668, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1668, "pid": 537, "tid": 537, "ts": 7429564847794.461,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847839.975, "dur": 0.349,
    "args": {
      "External id": 4524,      "cid": 370, "correlation": 1669
    }
  },
  {
    "ph": "f", "id": 1669, "pid": 537, "tid": 537, "ts": 7429564847839.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847844.377, "dur": 0.157,
    "args": {
      "External id": 4524,      "cid": 370, "correlation": 1670
    }
  },
  {
    "ph": "f", "id": 1670, "pid": 537, "tid": 537, "ts": 7429564847844.377,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847846.602, "dur": 0.150,
    "args": {
      "External id": 4524,      "cid": 370, "correlation": 1671
    }
  },
  {
    "ph": "f", "id": 1671, "pid": 537, "tid": 537, "ts": 7429564847846.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847849.159, "dur": 4.796,
    "args": {
      "External id": 4524,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1672,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1672, "pid": 537, "tid": 537, "ts": 7429564847849.159,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847893.128, "dur": 0.172,
    "args": {
      "External id": 4537,      "cid": 370, "correlation": 1673
    }
  },
  {
    "ph": "f", "id": 1673, "pid": 537, "tid": 537, "ts": 7429564847893.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847896.146, "dur": 0.152,
    "args": {
      "External id": 4537,      "cid": 370, "correlation": 1674
    }
  },
  {
    "ph": "f", "id": 1674, "pid": 537, "tid": 537, "ts": 7429564847896.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847898.047, "dur": 0.162,
    "args": {
      "External id": 4537,      "cid": 370, "correlation": 1675
    }
  },
  {
    "ph": "f", "id": 1675, "pid": 537, "tid": 537, "ts": 7429564847898.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847902.857, "dur": 5.124,
    "args": {
      "External id": 4537,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1676,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1676, "pid": 537, "tid": 537, "ts": 7429564847902.857,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847908.406, "dur": 4.515,
    "args": {
      "External id": 4537,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1677,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1677, "pid": 537, "tid": 537, "ts": 7429564847908.406,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847948.025, "dur": 0.221,
    "args": {
      "External id": 4550,      "cid": 370, "correlation": 1678
    }
  },
  {
    "ph": "f", "id": 1678, "pid": 537, "tid": 537, "ts": 7429564847948.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847950.809, "dur": 0.149,
    "args": {
      "External id": 4550,      "cid": 370, "correlation": 1679
    }
  },
  {
    "ph": "f", "id": 1679, "pid": 537, "tid": 537, "ts": 7429564847950.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564847952.624, "dur": 0.141,
    "args": {
      "External id": 4550,      "cid": 370, "correlation": 1680
    }
  },
  {
    "ph": "f", "id": 1680, "pid": 537, "tid": 537, "ts": 7429564847952.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847956.097, "dur": 4.725,
    "args": {
      "External id": 4550,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1681,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1681, "pid": 537, "tid": 537, "ts": 7429564847956.097,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847961.282, "dur": 5.235,
    "args": {
      "External id": 4550,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1682,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1682, "pid": 537, "tid": 537, "ts": 7429564847961.282,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564847991.208, "dur": 4.383,
    "args": {
      "External id": 4560,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1683,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1683, "pid": 537, "tid": 537, "ts": 7429564847991.208,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848018.821, "dur": 4.461,
    "args": {
      "External id": 4565,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1684,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1684, "pid": 537, "tid": 537, "ts": 7429564848018.821,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848036.587, "dur": 6.700,
    "args": {
      "External id": 4566,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1685,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1685, "pid": 537, "tid": 537, "ts": 7429564848036.587,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848051.155, "dur": 5.634,
    "args": {
      "External id": 4567,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1686,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1686, "pid": 537, "tid": 537, "ts": 7429564848051.155,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848062.079, "dur": 4.318,
    "args": {
      "External id": 4568,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1687,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1687, "pid": 537, "tid": 537, "ts": 7429564848062.079,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848072.084, "dur": 5.023,
    "args": {
      "External id": 4569,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1688,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1688, "pid": 537, "tid": 537, "ts": 7429564848072.084,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848090.755, "dur": 3.798,
    "args": {
      "External id": 4574,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1689,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1689, "pid": 537, "tid": 537, "ts": 7429564848090.755,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848101.336, "dur": 5.998,
    "args": {
      "External id": 4575,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1690,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1690, "pid": 537, "tid": 537, "ts": 7429564848101.336,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848114.016, "dur": 3.191,
    "args": {
      "External id": 4576,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1691,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1691, "pid": 537, "tid": 537, "ts": 7429564848114.016,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848122.401, "dur": 3.179,
    "args": {
      "External id": 4577,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1692,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1692, "pid": 537, "tid": 537, "ts": 7429564848122.401,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848184.251, "dur": 4.821,
    "args": {
      "External id": 4594,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1693,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1693, "pid": 537, "tid": 537, "ts": 7429564848184.251,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848249.795, "dur": 4.694,
    "args": {
      "External id": 4612,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1694,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1694, "pid": 537, "tid": 537, "ts": 7429564848249.795,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848289.277, "dur": 0.312,
    "args": {
      "External id": 4625,      "cid": 370, "correlation": 1695
    }
  },
  {
    "ph": "f", "id": 1695, "pid": 537, "tid": 537, "ts": 7429564848289.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848293.527, "dur": 0.167,
    "args": {
      "External id": 4625,      "cid": 370, "correlation": 1696
    }
  },
  {
    "ph": "f", "id": 1696, "pid": 537, "tid": 537, "ts": 7429564848293.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848296.017, "dur": 0.154,
    "args": {
      "External id": 4625,      "cid": 370, "correlation": 1697
    }
  },
  {
    "ph": "f", "id": 1697, "pid": 537, "tid": 537, "ts": 7429564848296.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848298.586, "dur": 4.866,
    "args": {
      "External id": 4625,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1698,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1698, "pid": 537, "tid": 537, "ts": 7429564848298.586,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848314.790, "dur": 3.824,
    "args": {
      "External id": 4628,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1699,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1699, "pid": 537, "tid": 537, "ts": 7429564848314.790,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848332.544, "dur": 3.619,
    "args": {
      "External id": 4635,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1700,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1700, "pid": 537, "tid": 537, "ts": 7429564848332.544,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848347.231, "dur": 5.513,
    "args": {
      "External id": 4638,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1701,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1701, "pid": 537, "tid": 537, "ts": 7429564848347.231,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848391.863, "dur": 0.202,
    "args": {
      "External id": 4650,      "cid": 370, "correlation": 1702
    }
  },
  {
    "ph": "f", "id": 1702, "pid": 537, "tid": 537, "ts": 7429564848391.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848395.438, "dur": 0.141,
    "args": {
      "External id": 4650,      "cid": 370, "correlation": 1703
    }
  },
  {
    "ph": "f", "id": 1703, "pid": 537, "tid": 537, "ts": 7429564848395.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848398.197, "dur": 0.156,
    "args": {
      "External id": 4650,      "cid": 370, "correlation": 1704
    }
  },
  {
    "ph": "f", "id": 1704, "pid": 537, "tid": 537, "ts": 7429564848398.197,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848400.225, "dur": 4.644,
    "args": {
      "External id": 4650,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1705,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1705, "pid": 537, "tid": 537, "ts": 7429564848400.225,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848420.948, "dur": 5.062,
    "args": {
      "External id": 4659,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1706,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1706, "pid": 537, "tid": 537, "ts": 7429564848420.948,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848464.568, "dur": 0.268,
    "args": {
      "External id": 4669,      "cid": 370, "correlation": 1707
    }
  },
  {
    "ph": "f", "id": 1707, "pid": 537, "tid": 537, "ts": 7429564848464.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848467.851, "dur": 0.157,
    "args": {
      "External id": 4669,      "cid": 370, "correlation": 1708
    }
  },
  {
    "ph": "f", "id": 1708, "pid": 537, "tid": 537, "ts": 7429564848467.851,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848469.818, "dur": 0.286,
    "args": {
      "External id": 4669,      "cid": 370, "correlation": 1709
    }
  },
  {
    "ph": "f", "id": 1709, "pid": 537, "tid": 537, "ts": 7429564848469.818,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848471.754, "dur": 5.706,
    "args": {
      "External id": 4669,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1710,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1710, "pid": 537, "tid": 537, "ts": 7429564848471.754,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848489.314, "dur": 4.890,
    "args": {
      "External id": 4672,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1711,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1711, "pid": 537, "tid": 537, "ts": 7429564848489.314,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848511.455, "dur": 4.359,
    "args": {
      "External id": 4674,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1712,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1712, "pid": 537, "tid": 537, "ts": 7429564848511.455,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848526.495, "dur": 5.682,
    "args": {
      "External id": 4677,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1713,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1713, "pid": 537, "tid": 537, "ts": 7429564848526.495,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848540.102, "dur": 5.657,
    "args": {
      "External id": 4678,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1714,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1714, "pid": 537, "tid": 537, "ts": 7429564848540.102,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848551.997, "dur": 5.524,
    "args": {
      "External id": 4679,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1715,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1715, "pid": 537, "tid": 537, "ts": 7429564848551.997,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848563.741, "dur": 3.492,
    "args": {
      "External id": 4680,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1716,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1716, "pid": 537, "tid": 537, "ts": 7429564848563.741,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848576.112, "dur": 3.598,
    "args": {
      "External id": 4682,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1717,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1717, "pid": 537, "tid": 537, "ts": 7429564848576.112,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564848605.827, "dur": 0.327,
    "args": {
      "External id": 4690,      "cid": 290, "correlation": 1718
    }
  },
  {
    "ph": "f", "id": 1718, "pid": 537, "tid": 537, "ts": 7429564848605.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564848606.480, "dur": 52.965,
    "args": {
      "External id": 4690,            "bytes": 2097152, "cid": 108, "correlation": 1719, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1719, "pid": 537, "tid": 537, "ts": 7429564848606.480,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848673.760, "dur": 0.246,
    "args": {
      "External id": 4690,      "cid": 370, "correlation": 1720
    }
  },
  {
    "ph": "f", "id": 1720, "pid": 537, "tid": 537, "ts": 7429564848673.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848677.488, "dur": 0.151,
    "args": {
      "External id": 4690,      "cid": 370, "correlation": 1721
    }
  },
  {
    "ph": "f", "id": 1721, "pid": 537, "tid": 537, "ts": 7429564848677.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848680.369, "dur": 0.153,
    "args": {
      "External id": 4690,      "cid": 370, "correlation": 1722
    }
  },
  {
    "ph": "f", "id": 1722, "pid": 537, "tid": 537, "ts": 7429564848680.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848688.422, "dur": 5.037,
    "args": {
      "External id": 4690,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1723,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1723, "pid": 537, "tid": 537, "ts": 7429564848688.422,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848694.035, "dur": 4.754,
    "args": {
      "External id": 4690,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1724,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1724, "pid": 537, "tid": 537, "ts": 7429564848694.035,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848718.072, "dur": 4.469,
    "args": {
      "External id": 4693,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1725,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1725, "pid": 537, "tid": 537, "ts": 7429564848718.072,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564848740.932, "dur": 0.179,
    "args": {
      "External id": 4701,      "cid": 290, "correlation": 1726
    }
  },
  {
    "ph": "f", "id": 1726, "pid": 537, "tid": 537, "ts": 7429564848740.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564848741.403, "dur": 30.509,
    "args": {
      "External id": 4701,            "bytes": 2097152, "cid": 108, "correlation": 1727, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1727, "pid": 537, "tid": 537, "ts": 7429564848741.403,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848782.981, "dur": 0.198,
    "args": {
      "External id": 4701,      "cid": 370, "correlation": 1728
    }
  },
  {
    "ph": "f", "id": 1728, "pid": 537, "tid": 537, "ts": 7429564848782.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848786.871, "dur": 0.140,
    "args": {
      "External id": 4701,      "cid": 370, "correlation": 1729
    }
  },
  {
    "ph": "f", "id": 1729, "pid": 537, "tid": 537, "ts": 7429564848786.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848789.474, "dur": 0.162,
    "args": {
      "External id": 4701,      "cid": 370, "correlation": 1730
    }
  },
  {
    "ph": "f", "id": 1730, "pid": 537, "tid": 537, "ts": 7429564848789.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848794.552, "dur": 6.110,
    "args": {
      "External id": 4701,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1731,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1731, "pid": 537, "tid": 537, "ts": 7429564848794.552,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848801.234, "dur": 3.876,
    "args": {
      "External id": 4701,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1732,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1732, "pid": 537, "tid": 537, "ts": 7429564848801.234,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848815.765, "dur": 3.915,
    "args": {
      "External id": 4704,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1733,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1733, "pid": 537, "tid": 537, "ts": 7429564848815.765,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848848.479, "dur": 0.161,
    "args": {
      "External id": 4712,      "cid": 370, "correlation": 1734
    }
  },
  {
    "ph": "f", "id": 1734, "pid": 537, "tid": 537, "ts": 7429564848848.479,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848851.525, "dur": 0.140,
    "args": {
      "External id": 4712,      "cid": 370, "correlation": 1735
    }
  },
  {
    "ph": "f", "id": 1735, "pid": 537, "tid": 537, "ts": 7429564848851.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564848853.499, "dur": 0.141,
    "args": {
      "External id": 4712,      "cid": 370, "correlation": 1736
    }
  },
  {
    "ph": "f", "id": 1736, "pid": 537, "tid": 537, "ts": 7429564848853.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848857.355, "dur": 4.913,
    "args": {
      "External id": 4712,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1737,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1737, "pid": 537, "tid": 537, "ts": 7429564848857.355,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848862.734, "dur": 3.414,
    "args": {
      "External id": 4712,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1738,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1738, "pid": 537, "tid": 537, "ts": 7429564848862.734,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848876.011, "dur": 5.446,
    "args": {
      "External id": 4715,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1739,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1739, "pid": 537, "tid": 537, "ts": 7429564848876.011,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848910.661, "dur": 5.182,
    "args": {
      "External id": 4717,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1740,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1740, "pid": 537, "tid": 537, "ts": 7429564848910.661,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848925.790, "dur": 5.280,
    "args": {
      "External id": 4720,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1741,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1741, "pid": 537, "tid": 537, "ts": 7429564848925.790,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848939.157, "dur": 3.505,
    "args": {
      "External id": 4721,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1742,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1742, "pid": 537, "tid": 537, "ts": 7429564848939.157,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848947.546, "dur": 4.832,
    "args": {
      "External id": 4722,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1743,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1743, "pid": 537, "tid": 537, "ts": 7429564848947.546,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848959.333, "dur": 3.931,
    "args": {
      "External id": 4723,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1744,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1744, "pid": 537, "tid": 537, "ts": 7429564848959.333,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564848970.709, "dur": 6.425,
    "args": {
      "External id": 4725,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1745,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1745, "pid": 537, "tid": 537, "ts": 7429564848970.709,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849018.500, "dur": 0.226,
    "args": {
      "External id": 4733,      "cid": 370, "correlation": 1746
    }
  },
  {
    "ph": "f", "id": 1746, "pid": 537, "tid": 537, "ts": 7429564849018.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849021.927, "dur": 0.148,
    "args": {
      "External id": 4733,      "cid": 370, "correlation": 1747
    }
  },
  {
    "ph": "f", "id": 1747, "pid": 537, "tid": 537, "ts": 7429564849021.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849024.155, "dur": 0.158,
    "args": {
      "External id": 4733,      "cid": 370, "correlation": 1748
    }
  },
  {
    "ph": "f", "id": 1748, "pid": 537, "tid": 537, "ts": 7429564849024.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849026.278, "dur": 6.180,
    "args": {
      "External id": 4733,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1749,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1749, "pid": 537, "tid": 537, "ts": 7429564849026.278,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849073.418, "dur": 0.235,
    "args": {
      "External id": 4746,      "cid": 370, "correlation": 1750
    }
  },
  {
    "ph": "f", "id": 1750, "pid": 537, "tid": 537, "ts": 7429564849073.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849076.534, "dur": 0.141,
    "args": {
      "External id": 4746,      "cid": 370, "correlation": 1751
    }
  },
  {
    "ph": "f", "id": 1751, "pid": 537, "tid": 537, "ts": 7429564849076.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849078.570, "dur": 0.158,
    "args": {
      "External id": 4746,      "cid": 370, "correlation": 1752
    }
  },
  {
    "ph": "f", "id": 1752, "pid": 537, "tid": 537, "ts": 7429564849078.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849082.899, "dur": 6.096,
    "args": {
      "External id": 4746,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1753,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1753, "pid": 537, "tid": 537, "ts": 7429564849082.899,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849089.646, "dur": 3.795,
    "args": {
      "External id": 4746,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1754,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1754, "pid": 537, "tid": 537, "ts": 7429564849089.646,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849128.326, "dur": 0.160,
    "args": {
      "External id": 4759,      "cid": 370, "correlation": 1755
    }
  },
  {
    "ph": "f", "id": 1755, "pid": 537, "tid": 537, "ts": 7429564849128.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849131.154, "dur": 0.142,
    "args": {
      "External id": 4759,      "cid": 370, "correlation": 1756
    }
  },
  {
    "ph": "f", "id": 1756, "pid": 537, "tid": 537, "ts": 7429564849131.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849132.905, "dur": 0.141,
    "args": {
      "External id": 4759,      "cid": 370, "correlation": 1757
    }
  },
  {
    "ph": "f", "id": 1757, "pid": 537, "tid": 537, "ts": 7429564849132.905,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849136.428, "dur": 4.066,
    "args": {
      "External id": 4759,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1758,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1758, "pid": 537, "tid": 537, "ts": 7429564849136.428,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849140.807, "dur": 4.234,
    "args": {
      "External id": 4759,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1759,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1759, "pid": 537, "tid": 537, "ts": 7429564849140.807,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849168.442, "dur": 4.863,
    "args": {
      "External id": 4769,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1760,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1760, "pid": 537, "tid": 537, "ts": 7429564849168.442,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849196.398, "dur": 6.792,
    "args": {
      "External id": 4774,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1761,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1761, "pid": 537, "tid": 537, "ts": 7429564849196.398,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849217.935, "dur": 7.530,
    "args": {
      "External id": 4775,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1762,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1762, "pid": 537, "tid": 537, "ts": 7429564849217.935,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564849232.898, "dur": 0.346,
    "args": {
      "External id": 4776,      "cid": 290, "correlation": 1763
    }
  },
  {
    "ph": "f", "id": 1763, "pid": 537, "tid": 537, "ts": 7429564849232.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564849233.800, "dur": 39.238,
    "args": {
      "External id": 4776,            "bytes": 2097152, "cid": 108, "correlation": 1764, "ptr": "0x2"
    }
  },
  {
    "ph": "s", "id": 1764, "pid": 537, "tid": 537, "ts": 7429564849233.800,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849275.293, "dur": 3.841,
    "args": {
      "External id": 4776,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1765,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1765, "pid": 537, "tid": 537, "ts": 7429564849275.293,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849285.957, "dur": 5.657,
    "args": {
      "External id": 4777,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1766,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1766, "pid": 537, "tid": 537, "ts": 7429564849285.957,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849299.061, "dur": 4.108,
    "args": {
      "External id": 4778,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1767,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1767, "pid": 537, "tid": 537, "ts": 7429564849299.061,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849317.870, "dur": 3.464,
    "args": {
      "External id": 4783,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1768,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1768, "pid": 537, "tid": 537, "ts": 7429564849317.870,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849329.625, "dur": 6.458,
    "args": {
      "External id": 4784,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1769,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1769, "pid": 537, "tid": 537, "ts": 7429564849329.625,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849342.345, "dur": 3.530,
    "args": {
      "External id": 4785,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1770,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1770, "pid": 537, "tid": 537, "ts": 7429564849342.345,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849350.399, "dur": 3.570,
    "args": {
      "External id": 4786,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1771,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1771, "pid": 537, "tid": 537, "ts": 7429564849350.399,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849417.355, "dur": 4.892,
    "args": {
      "External id": 4803,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1772,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1772, "pid": 537, "tid": 537, "ts": 7429564849417.355,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849453.970, "dur": 3.760,
    "args": {
      "External id": 4821,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1773,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1773, "pid": 537, "tid": 537, "ts": 7429564849453.970,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849494.354, "dur": 0.326,
    "args": {
      "External id": 4834,      "cid": 370, "correlation": 1774
    }
  },
  {
    "ph": "f", "id": 1774, "pid": 537, "tid": 537, "ts": 7429564849494.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849499.172, "dur": 0.144,
    "args": {
      "External id": 4834,      "cid": 370, "correlation": 1775
    }
  },
  {
    "ph": "f", "id": 1775, "pid": 537, "tid": 537, "ts": 7429564849499.172,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849501.681, "dur": 0.185,
    "args": {
      "External id": 4834,      "cid": 370, "correlation": 1776
    }
  },
  {
    "ph": "f", "id": 1776, "pid": 537, "tid": 537, "ts": 7429564849501.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849504.814, "dur": 5.088,
    "args": {
      "External id": 4834,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1777,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1777, "pid": 537, "tid": 537, "ts": 7429564849504.814,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849522.247, "dur": 4.244,
    "args": {
      "External id": 4837,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1778,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1778, "pid": 537, "tid": 537, "ts": 7429564849522.247,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849540.394, "dur": 4.126,
    "args": {
      "External id": 4844,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1779,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1779, "pid": 537, "tid": 537, "ts": 7429564849540.394,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849556.324, "dur": 5.448,
    "args": {
      "External id": 4847,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1780,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1780, "pid": 537, "tid": 537, "ts": 7429564849556.324,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849604.772, "dur": 0.215,
    "args": {
      "External id": 4859,      "cid": 370, "correlation": 1781
    }
  },
  {
    "ph": "f", "id": 1781, "pid": 537, "tid": 537, "ts": 7429564849604.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849608.299, "dur": 0.139,
    "args": {
      "External id": 4859,      "cid": 370, "correlation": 1782
    }
  },
  {
    "ph": "f", "id": 1782, "pid": 537, "tid": 537, "ts": 7429564849608.299,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849610.997, "dur": 0.152,
    "args": {
      "External id": 4859,      "cid": 370, "correlation": 1783
    }
  },
  {
    "ph": "f", "id": 1783, "pid": 537, "tid": 537, "ts": 7429564849610.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849613.250, "dur": 4.908,
    "args": {
      "External id": 4859,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1784,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1784, "pid": 537, "tid": 537, "ts": 7429564849613.250,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849635.125, "dur": 4.545,
    "args": {
      "External id": 4868,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1785,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1785, "pid": 537, "tid": 537, "ts": 7429564849635.125,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849682.587, "dur": 0.297,
    "args": {
      "External id": 4878,      "cid": 370, "correlation": 1786
    }
  },
  {
    "ph": "f", "id": 1786, "pid": 537, "tid": 537, "ts": 7429564849682.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849685.921, "dur": 0.146,
    "args": {
      "External id": 4878,      "cid": 370, "correlation": 1787
    }
  },
  {
    "ph": "f", "id": 1787, "pid": 537, "tid": 537, "ts": 7429564849685.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849688.384, "dur": 0.155,
    "args": {
      "External id": 4878,      "cid": 370, "correlation": 1788
    }
  },
  {
    "ph": "f", "id": 1788, "pid": 537, "tid": 537, "ts": 7429564849688.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849690.166, "dur": 4.796,
    "args": {
      "External id": 4878,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1789,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1789, "pid": 537, "tid": 537, "ts": 7429564849690.166,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849707.554, "dur": 4.580,
    "args": {
      "External id": 4881,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1790,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1790, "pid": 537, "tid": 537, "ts": 7429564849707.554,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849729.234, "dur": 4.222,
    "args": {
      "External id": 4883,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1791,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1791, "pid": 537, "tid": 537, "ts": 7429564849729.234,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849743.585, "dur": 4.769,
    "args": {
      "External id": 4886,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1792,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1792, "pid": 537, "tid": 537, "ts": 7429564849743.585,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849756.726, "dur": 4.284,
    "args": {
      "External id": 4887,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1793,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1793, "pid": 537, "tid": 537, "ts": 7429564849756.726,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849768.669, "dur": 4.134,
    "args": {
      "External id": 4888,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1794,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1794, "pid": 537, "tid": 537, "ts": 7429564849768.669,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849779.097, "dur": 5.130,
    "args": {
      "External id": 4889,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1795,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1795, "pid": 537, "tid": 537, "ts": 7429564849779.097,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849793.368, "dur": 5.342,
    "args": {
      "External id": 4891,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1796,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1796, "pid": 537, "tid": 537, "ts": 7429564849793.368,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849834.580, "dur": 0.245,
    "args": {
      "External id": 4899,      "cid": 370, "correlation": 1797
    }
  },
  {
    "ph": "f", "id": 1797, "pid": 537, "tid": 537, "ts": 7429564849834.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849837.995, "dur": 0.155,
    "args": {
      "External id": 4899,      "cid": 370, "correlation": 1798
    }
  },
  {
    "ph": "f", "id": 1798, "pid": 537, "tid": 537, "ts": 7429564849837.995,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849840.128, "dur": 0.156,
    "args": {
      "External id": 4899,      "cid": 370, "correlation": 1799
    }
  },
  {
    "ph": "f", "id": 1799, "pid": 537, "tid": 537, "ts": 7429564849840.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849846.939, "dur": 6.148,
    "args": {
      "External id": 4899,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1800,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1800, "pid": 537, "tid": 537, "ts": 7429564849846.939,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849853.444, "dur": 4.529,
    "args": {
      "External id": 4899,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1801,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1801, "pid": 537, "tid": 537, "ts": 7429564849853.444,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564849874.032, "dur": 0.379,
    "args": {
      "External id": 4902,      "cid": 290, "correlation": 1802
    }
  },
  {
    "ph": "f", "id": 1802, "pid": 537, "tid": 537, "ts": 7429564849874.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564849874.848, "dur": 45.351,
    "args": {
      "External id": 4902,            "bytes": 2097152, "cid": 108, "correlation": 1803, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 1803, "pid": 537, "tid": 537, "ts": 7429564849874.848,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849923.551, "dur": 4.683,
    "args": {
      "External id": 4902,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1804,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1804, "pid": 537, "tid": 537, "ts": 7429564849923.551,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849961.899, "dur": 0.196,
    "args": {
      "External id": 4910,      "cid": 370, "correlation": 1805
    }
  },
  {
    "ph": "f", "id": 1805, "pid": 537, "tid": 537, "ts": 7429564849961.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849965.246, "dur": 0.141,
    "args": {
      "External id": 4910,      "cid": 370, "correlation": 1806
    }
  },
  {
    "ph": "f", "id": 1806, "pid": 537, "tid": 537, "ts": 7429564849965.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564849967.370, "dur": 0.160,
    "args": {
      "External id": 4910,      "cid": 370, "correlation": 1807
    }
  },
  {
    "ph": "f", "id": 1807, "pid": 537, "tid": 537, "ts": 7429564849967.370,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849972.131, "dur": 6.517,
    "args": {
      "External id": 4910,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1808,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1808, "pid": 537, "tid": 537, "ts": 7429564849972.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564849978.999, "dur": 3.767,
    "args": {
      "External id": 4910,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1809,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1809, "pid": 537, "tid": 537, "ts": 7429564849978.999,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564849991.119, "dur": 0.203,
    "args": {
      "External id": 4913,      "cid": 290, "correlation": 1810
    }
  },
  {
    "ph": "f", "id": 1810, "pid": 537, "tid": 537, "ts": 7429564849991.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564849991.638, "dur": 30.061,
    "args": {
      "External id": 4913,            "bytes": 2097152, "cid": 108, "correlation": 1811, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 1811, "pid": 537, "tid": 537, "ts": 7429564849991.638,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850024.092, "dur": 5.203,
    "args": {
      "External id": 4913,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1812,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1812, "pid": 537, "tid": 537, "ts": 7429564850024.092,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850056.243, "dur": 0.206,
    "args": {
      "External id": 4921,      "cid": 370, "correlation": 1813
    }
  },
  {
    "ph": "f", "id": 1813, "pid": 537, "tid": 537, "ts": 7429564850056.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850059.748, "dur": 0.144,
    "args": {
      "External id": 4921,      "cid": 370, "correlation": 1814
    }
  },
  {
    "ph": "f", "id": 1814, "pid": 537, "tid": 537, "ts": 7429564850059.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850062.101, "dur": 0.145,
    "args": {
      "External id": 4921,      "cid": 370, "correlation": 1815
    }
  },
  {
    "ph": "f", "id": 1815, "pid": 537, "tid": 537, "ts": 7429564850062.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850066.474, "dur": 21.907,
    "args": {
      "External id": 4921,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1816,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1816, "pid": 537, "tid": 537, "ts": 7429564850066.474,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850089.125, "dur": 3.498,
    "args": {
      "External id": 4921,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1817,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1817, "pid": 537, "tid": 537, "ts": 7429564850089.125,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850109.374, "dur": 5.671,
    "args": {
      "External id": 4924,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1818,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1818, "pid": 537, "tid": 537, "ts": 7429564850109.374,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850145.481, "dur": 4.547,
    "args": {
      "External id": 4926,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1819,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1819, "pid": 537, "tid": 537, "ts": 7429564850145.481,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850160.670, "dur": 5.501,
    "args": {
      "External id": 4929,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1820,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1820, "pid": 537, "tid": 537, "ts": 7429564850160.670,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850176.046, "dur": 4.747,
    "args": {
      "External id": 4930,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1821,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1821, "pid": 537, "tid": 537, "ts": 7429564850176.046,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850186.457, "dur": 4.809,
    "args": {
      "External id": 4931,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1822,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1822, "pid": 537, "tid": 537, "ts": 7429564850186.457,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850197.938, "dur": 3.835,
    "args": {
      "External id": 4932,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1823,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1823, "pid": 537, "tid": 537, "ts": 7429564850197.938,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850209.963, "dur": 3.787,
    "args": {
      "External id": 4934,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1824,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1824, "pid": 537, "tid": 537, "ts": 7429564850209.963,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850256.659, "dur": 0.308,
    "args": {
      "External id": 4942,      "cid": 370, "correlation": 1825
    }
  },
  {
    "ph": "f", "id": 1825, "pid": 537, "tid": 537, "ts": 7429564850256.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850260.644, "dur": 0.165,
    "args": {
      "External id": 4942,      "cid": 370, "correlation": 1826
    }
  },
  {
    "ph": "f", "id": 1826, "pid": 537, "tid": 537, "ts": 7429564850260.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850263.080, "dur": 0.167,
    "args": {
      "External id": 4942,      "cid": 370, "correlation": 1827
    }
  },
  {
    "ph": "f", "id": 1827, "pid": 537, "tid": 537, "ts": 7429564850263.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850265.793, "dur": 4.676,
    "args": {
      "External id": 4942,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1828,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1828, "pid": 537, "tid": 537, "ts": 7429564850265.793,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850308.049, "dur": 0.196,
    "args": {
      "External id": 4955,      "cid": 370, "correlation": 1829
    }
  },
  {
    "ph": "f", "id": 1829, "pid": 537, "tid": 537, "ts": 7429564850308.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850311.507, "dur": 0.155,
    "args": {
      "External id": 4955,      "cid": 370, "correlation": 1830
    }
  },
  {
    "ph": "f", "id": 1830, "pid": 537, "tid": 537, "ts": 7429564850311.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850313.519, "dur": 0.169,
    "args": {
      "External id": 4955,      "cid": 370, "correlation": 1831
    }
  },
  {
    "ph": "f", "id": 1831, "pid": 537, "tid": 537, "ts": 7429564850313.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850318.488, "dur": 4.957,
    "args": {
      "External id": 4955,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1832,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1832, "pid": 537, "tid": 537, "ts": 7429564850318.488,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850323.854, "dur": 4.959,
    "args": {
      "External id": 4955,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1833,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1833, "pid": 537, "tid": 537, "ts": 7429564850323.854,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850363.510, "dur": 0.199,
    "args": {
      "External id": 4968,      "cid": 370, "correlation": 1834
    }
  },
  {
    "ph": "f", "id": 1834, "pid": 537, "tid": 537, "ts": 7429564850363.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850366.139, "dur": 0.143,
    "args": {
      "External id": 4968,      "cid": 370, "correlation": 1835
    }
  },
  {
    "ph": "f", "id": 1835, "pid": 537, "tid": 537, "ts": 7429564850366.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850368.198, "dur": 0.149,
    "args": {
      "External id": 4968,      "cid": 370, "correlation": 1836
    }
  },
  {
    "ph": "f", "id": 1836, "pid": 537, "tid": 537, "ts": 7429564850368.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850371.614, "dur": 4.644,
    "args": {
      "External id": 4968,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1837,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1837, "pid": 537, "tid": 537, "ts": 7429564850371.614,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850376.723, "dur": 3.350,
    "args": {
      "External id": 4968,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1838,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1838, "pid": 537, "tid": 537, "ts": 7429564850376.723,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850403.211, "dur": 3.626,
    "args": {
      "External id": 4978,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1839,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1839, "pid": 537, "tid": 537, "ts": 7429564850403.211,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850429.510, "dur": 4.910,
    "args": {
      "External id": 4983,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1840,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1840, "pid": 537, "tid": 537, "ts": 7429564850429.510,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850448.444, "dur": 7.880,
    "args": {
      "External id": 4984,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1841,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1841, "pid": 537, "tid": 537, "ts": 7429564850448.444,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850463.557, "dur": 5.101,
    "args": {
      "External id": 4985,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1842,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1842, "pid": 537, "tid": 537, "ts": 7429564850463.557,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850473.848, "dur": 4.593,
    "args": {
      "External id": 4986,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1843,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1843, "pid": 537, "tid": 537, "ts": 7429564850473.848,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850484.906, "dur": 3.320,
    "args": {
      "External id": 4987,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1844,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1844, "pid": 537, "tid": 537, "ts": 7429564850484.906,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850500.838, "dur": 5.152,
    "args": {
      "External id": 4992,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1845,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1845, "pid": 537, "tid": 537, "ts": 7429564850500.838,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850512.941, "dur": 4.389,
    "args": {
      "External id": 4993,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1846,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1846, "pid": 537, "tid": 537, "ts": 7429564850512.941,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850522.805, "dur": 3.410,
    "args": {
      "External id": 4994,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1847,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1847, "pid": 537, "tid": 537, "ts": 7429564850522.805,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850531.505, "dur": 5.224,
    "args": {
      "External id": 4995,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1848,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1848, "pid": 537, "tid": 537, "ts": 7429564850531.505,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850597.391, "dur": 5.947,
    "args": {
      "External id": 5012,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1849,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1849, "pid": 537, "tid": 537, "ts": 7429564850597.391,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850635.017, "dur": 3.517,
    "args": {
      "External id": 5030,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1850,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1850, "pid": 537, "tid": 537, "ts": 7429564850635.017,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850674.423, "dur": 0.295,
    "args": {
      "External id": 5043,      "cid": 370, "correlation": 1851
    }
  },
  {
    "ph": "f", "id": 1851, "pid": 537, "tid": 537, "ts": 7429564850674.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850678.807, "dur": 0.143,
    "args": {
      "External id": 5043,      "cid": 370, "correlation": 1852
    }
  },
  {
    "ph": "f", "id": 1852, "pid": 537, "tid": 537, "ts": 7429564850678.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850681.065, "dur": 0.168,
    "args": {
      "External id": 5043,      "cid": 370, "correlation": 1853
    }
  },
  {
    "ph": "f", "id": 1853, "pid": 537, "tid": 537, "ts": 7429564850681.065,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850683.775, "dur": 4.661,
    "args": {
      "External id": 5043,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1854,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1854, "pid": 537, "tid": 537, "ts": 7429564850683.775,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850700.456, "dur": 4.815,
    "args": {
      "External id": 5046,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1855,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1855, "pid": 537, "tid": 537, "ts": 7429564850700.456,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850719.636, "dur": 4.761,
    "args": {
      "External id": 5053,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1856,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1856, "pid": 537, "tid": 537, "ts": 7429564850719.636,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850736.481, "dur": 6.708,
    "args": {
      "External id": 5056,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1857,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1857, "pid": 537, "tid": 537, "ts": 7429564850736.481,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850785.547, "dur": 0.181,
    "args": {
      "External id": 5068,      "cid": 370, "correlation": 1858
    }
  },
  {
    "ph": "f", "id": 1858, "pid": 537, "tid": 537, "ts": 7429564850785.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850789.231, "dur": 0.147,
    "args": {
      "External id": 5068,      "cid": 370, "correlation": 1859
    }
  },
  {
    "ph": "f", "id": 1859, "pid": 537, "tid": 537, "ts": 7429564850789.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850792.040, "dur": 0.163,
    "args": {
      "External id": 5068,      "cid": 370, "correlation": 1860
    }
  },
  {
    "ph": "f", "id": 1860, "pid": 537, "tid": 537, "ts": 7429564850792.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850794.200, "dur": 5.968,
    "args": {
      "External id": 5068,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1861,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1861, "pid": 537, "tid": 537, "ts": 7429564850794.200,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850816.345, "dur": 3.755,
    "args": {
      "External id": 5077,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1862,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1862, "pid": 537, "tid": 537, "ts": 7429564850816.345,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850863.067, "dur": 0.176,
    "args": {
      "External id": 5087,      "cid": 370, "correlation": 1863
    }
  },
  {
    "ph": "f", "id": 1863, "pid": 537, "tid": 537, "ts": 7429564850863.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850866.050, "dur": 0.160,
    "args": {
      "External id": 5087,      "cid": 370, "correlation": 1864
    }
  },
  {
    "ph": "f", "id": 1864, "pid": 537, "tid": 537, "ts": 7429564850866.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564850868.497, "dur": 0.166,
    "args": {
      "External id": 5087,      "cid": 370, "correlation": 1865
    }
  },
  {
    "ph": "f", "id": 1865, "pid": 537, "tid": 537, "ts": 7429564850868.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850870.418, "dur": 4.133,
    "args": {
      "External id": 5087,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1866,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1866, "pid": 537, "tid": 537, "ts": 7429564850870.418,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850886.748, "dur": 4.294,
    "args": {
      "External id": 5090,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1867,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1867, "pid": 537, "tid": 537, "ts": 7429564850886.748,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850908.370, "dur": 4.798,
    "args": {
      "External id": 5092,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1868,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1868, "pid": 537, "tid": 537, "ts": 7429564850908.370,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850922.894, "dur": 4.791,
    "args": {
      "External id": 5095,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1869,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1869, "pid": 537, "tid": 537, "ts": 7429564850922.894,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850935.658, "dur": 4.211,
    "args": {
      "External id": 5096,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1870,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1870, "pid": 537, "tid": 537, "ts": 7429564850935.658,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850946.517, "dur": 3.849,
    "args": {
      "External id": 5097,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1871,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1871, "pid": 537, "tid": 537, "ts": 7429564850946.517,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850956.512, "dur": 3.851,
    "args": {
      "External id": 5098,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1872,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1872, "pid": 537, "tid": 537, "ts": 7429564850956.512,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564850969.287, "dur": 3.220,
    "args": {
      "External id": 5100,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1873,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1873, "pid": 537, "tid": 537, "ts": 7429564850969.287,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564851000.244, "dur": 0.426,
    "args": {
      "External id": 5108,      "cid": 290, "correlation": 1874
    }
  },
  {
    "ph": "f", "id": 1874, "pid": 537, "tid": 537, "ts": 7429564851000.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564851001.158, "dur": 58.079,
    "args": {
      "External id": 5108,            "bytes": 2097152, "cid": 108, "correlation": 1875, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1875, "pid": 537, "tid": 537, "ts": 7429564851001.158,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851074.599, "dur": 0.283,
    "args": {
      "External id": 5108,      "cid": 370, "correlation": 1876
    }
  },
  {
    "ph": "f", "id": 1876, "pid": 537, "tid": 537, "ts": 7429564851074.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851079.047, "dur": 0.159,
    "args": {
      "External id": 5108,      "cid": 370, "correlation": 1877
    }
  },
  {
    "ph": "f", "id": 1877, "pid": 537, "tid": 537, "ts": 7429564851079.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851081.822, "dur": 0.146,
    "args": {
      "External id": 5108,      "cid": 370, "correlation": 1878
    }
  },
  {
    "ph": "f", "id": 1878, "pid": 537, "tid": 537, "ts": 7429564851081.822,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851090.448, "dur": 5.484,
    "args": {
      "External id": 5108,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1879,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1879, "pid": 537, "tid": 537, "ts": 7429564851090.448,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851096.315, "dur": 4.883,
    "args": {
      "External id": 5108,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1880,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1880, "pid": 537, "tid": 537, "ts": 7429564851096.315,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851122.371, "dur": 5.312,
    "args": {
      "External id": 5111,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1881,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1881, "pid": 537, "tid": 537, "ts": 7429564851122.371,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564851148.835, "dur": 0.315,
    "args": {
      "External id": 5119,      "cid": 290, "correlation": 1882
    }
  },
  {
    "ph": "f", "id": 1882, "pid": 537, "tid": 537, "ts": 7429564851148.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564851149.532, "dur": 30.112,
    "args": {
      "External id": 5119,            "bytes": 2097152, "cid": 108, "correlation": 1883, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 1883, "pid": 537, "tid": 537, "ts": 7429564851149.532,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851192.289, "dur": 0.213,
    "args": {
      "External id": 5119,      "cid": 370, "correlation": 1884
    }
  },
  {
    "ph": "f", "id": 1884, "pid": 537, "tid": 537, "ts": 7429564851192.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851195.496, "dur": 0.153,
    "args": {
      "External id": 5119,      "cid": 370, "correlation": 1885
    }
  },
  {
    "ph": "f", "id": 1885, "pid": 537, "tid": 537, "ts": 7429564851195.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851198.236, "dur": 0.159,
    "args": {
      "External id": 5119,      "cid": 370, "correlation": 1886
    }
  },
  {
    "ph": "f", "id": 1886, "pid": 537, "tid": 537, "ts": 7429564851198.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851203.343, "dur": 5.100,
    "args": {
      "External id": 5119,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1887,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1887, "pid": 537, "tid": 537, "ts": 7429564851203.343,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851208.785, "dur": 4.389,
    "args": {
      "External id": 5119,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1888,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1888, "pid": 537, "tid": 537, "ts": 7429564851208.785,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851255.640, "dur": 4.379,
    "args": {
      "External id": 5122,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1889,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1889, "pid": 537, "tid": 537, "ts": 7429564851255.640,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851285.672, "dur": 0.183,
    "args": {
      "External id": 5130,      "cid": 370, "correlation": 1890
    }
  },
  {
    "ph": "f", "id": 1890, "pid": 537, "tid": 537, "ts": 7429564851285.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851289.220, "dur": 0.140,
    "args": {
      "External id": 5130,      "cid": 370, "correlation": 1891
    }
  },
  {
    "ph": "f", "id": 1891, "pid": 537, "tid": 537, "ts": 7429564851289.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851291.488, "dur": 0.157,
    "args": {
      "External id": 5130,      "cid": 370, "correlation": 1892
    }
  },
  {
    "ph": "f", "id": 1892, "pid": 537, "tid": 537, "ts": 7429564851291.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851295.177, "dur": 5.793,
    "args": {
      "External id": 5130,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1893,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1893, "pid": 537, "tid": 537, "ts": 7429564851295.177,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851301.474, "dur": 4.997,
    "args": {
      "External id": 5130,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1894,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1894, "pid": 537, "tid": 537, "ts": 7429564851301.474,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851316.549, "dur": 4.556,
    "args": {
      "External id": 5133,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1895,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1895, "pid": 537, "tid": 537, "ts": 7429564851316.549,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851349.668, "dur": 4.114,
    "args": {
      "External id": 5135,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1896,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1896, "pid": 537, "tid": 537, "ts": 7429564851349.668,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851363.926, "dur": 5.901,
    "args": {
      "External id": 5138,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1897,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1897, "pid": 537, "tid": 537, "ts": 7429564851363.926,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851378.015, "dur": 3.615,
    "args": {
      "External id": 5139,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1898,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1898, "pid": 537, "tid": 537, "ts": 7429564851378.015,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851386.526, "dur": 7.028,
    "args": {
      "External id": 5140,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1899,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1899, "pid": 537, "tid": 537, "ts": 7429564851386.526,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851401.077, "dur": 4.119,
    "args": {
      "External id": 5141,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1900,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1900, "pid": 537, "tid": 537, "ts": 7429564851401.077,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851412.563, "dur": 5.121,
    "args": {
      "External id": 5143,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1901,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1901, "pid": 537, "tid": 537, "ts": 7429564851412.563,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851458.130, "dur": 0.316,
    "args": {
      "External id": 5151,      "cid": 370, "correlation": 1902
    }
  },
  {
    "ph": "f", "id": 1902, "pid": 537, "tid": 537, "ts": 7429564851458.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851461.677, "dur": 0.158,
    "args": {
      "External id": 5151,      "cid": 370, "correlation": 1903
    }
  },
  {
    "ph": "f", "id": 1903, "pid": 537, "tid": 537, "ts": 7429564851461.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851463.705, "dur": 0.170,
    "args": {
      "External id": 5151,      "cid": 370, "correlation": 1904
    }
  },
  {
    "ph": "f", "id": 1904, "pid": 537, "tid": 537, "ts": 7429564851463.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851465.832, "dur": 4.548,
    "args": {
      "External id": 5151,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1905,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1905, "pid": 537, "tid": 537, "ts": 7429564851465.832,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851509.485, "dur": 0.199,
    "args": {
      "External id": 5164,      "cid": 370, "correlation": 1906
    }
  },
  {
    "ph": "f", "id": 1906, "pid": 537, "tid": 537, "ts": 7429564851509.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851512.478, "dur": 0.155,
    "args": {
      "External id": 5164,      "cid": 370, "correlation": 1907
    }
  },
  {
    "ph": "f", "id": 1907, "pid": 537, "tid": 537, "ts": 7429564851512.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851514.464, "dur": 0.163,
    "args": {
      "External id": 5164,      "cid": 370, "correlation": 1908
    }
  },
  {
    "ph": "f", "id": 1908, "pid": 537, "tid": 537, "ts": 7429564851514.464,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851518.613, "dur": 4.587,
    "args": {
      "External id": 5164,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1909,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1909, "pid": 537, "tid": 537, "ts": 7429564851518.613,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851523.776, "dur": 5.225,
    "args": {
      "External id": 5164,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1910,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1910, "pid": 537, "tid": 537, "ts": 7429564851523.776,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851562.420, "dur": 0.172,
    "args": {
      "External id": 5177,      "cid": 370, "correlation": 1911
    }
  },
  {
    "ph": "f", "id": 1911, "pid": 537, "tid": 537, "ts": 7429564851562.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851565.114, "dur": 0.144,
    "args": {
      "External id": 5177,      "cid": 370, "correlation": 1912
    }
  },
  {
    "ph": "f", "id": 1912, "pid": 537, "tid": 537, "ts": 7429564851565.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851566.937, "dur": 0.152,
    "args": {
      "External id": 5177,      "cid": 370, "correlation": 1913
    }
  },
  {
    "ph": "f", "id": 1913, "pid": 537, "tid": 537, "ts": 7429564851566.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851570.497, "dur": 5.398,
    "args": {
      "External id": 5177,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1914,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1914, "pid": 537, "tid": 537, "ts": 7429564851570.497,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851576.194, "dur": 3.851,
    "args": {
      "External id": 5177,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1915,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1915, "pid": 537, "tid": 537, "ts": 7429564851576.194,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564851601.776, "dur": 0.298,
    "args": {
      "External id": 5187,      "cid": 290, "correlation": 1916
    }
  },
  {
    "ph": "f", "id": 1916, "pid": 537, "tid": 537, "ts": 7429564851601.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564851602.469, "dur": 35.919,
    "args": {
      "External id": 5187,            "bytes": 2097152, "cid": 108, "correlation": 1917, "ptr": "0x7fee32880142"
    }
  },
  {
    "ph": "s", "id": 1917, "pid": 537, "tid": 537, "ts": 7429564851602.469,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851641.653, "dur": 5.171,
    "args": {
      "External id": 5187,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1918,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1918, "pid": 537, "tid": 537, "ts": 7429564851641.653,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851673.095, "dur": 6.003,
    "args": {
      "External id": 5192,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1919,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1919, "pid": 537, "tid": 537, "ts": 7429564851673.095,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851694.068, "dur": 7.609,
    "args": {
      "External id": 5193,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1920,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1920, "pid": 537, "tid": 537, "ts": 7429564851694.068,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851710.752, "dur": 4.051,
    "args": {
      "External id": 5194,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1921,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1921, "pid": 537, "tid": 537, "ts": 7429564851710.752,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851720.584, "dur": 4.839,
    "args": {
      "External id": 5195,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1922,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1922, "pid": 537, "tid": 537, "ts": 7429564851720.584,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851731.769, "dur": 4.586,
    "args": {
      "External id": 5196,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1923,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1923, "pid": 537, "tid": 537, "ts": 7429564851731.769,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851749.424, "dur": 3.695,
    "args": {
      "External id": 5201,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1924,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1924, "pid": 537, "tid": 537, "ts": 7429564851749.424,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851759.413, "dur": 4.339,
    "args": {
      "External id": 5202,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1925,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1925, "pid": 537, "tid": 537, "ts": 7429564851759.413,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851769.884, "dur": 3.436,
    "args": {
      "External id": 5203,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1926,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1926, "pid": 537, "tid": 537, "ts": 7429564851769.884,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851777.578, "dur": 3.787,
    "args": {
      "External id": 5204,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1927,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1927, "pid": 537, "tid": 537, "ts": 7429564851777.578,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851841.562, "dur": 5.770,
    "args": {
      "External id": 5221,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1928,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1928, "pid": 537, "tid": 537, "ts": 7429564851841.562,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851877.227, "dur": 4.569,
    "args": {
      "External id": 5239,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1929,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1929, "pid": 537, "tid": 537, "ts": 7429564851877.227,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851917.246, "dur": 0.308,
    "args": {
      "External id": 5252,      "cid": 370, "correlation": 1930
    }
  },
  {
    "ph": "f", "id": 1930, "pid": 537, "tid": 537, "ts": 7429564851917.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851921.701, "dur": 0.158,
    "args": {
      "External id": 5252,      "cid": 370, "correlation": 1931
    }
  },
  {
    "ph": "f", "id": 1931, "pid": 537, "tid": 537, "ts": 7429564851921.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564851924.112, "dur": 0.157,
    "args": {
      "External id": 5252,      "cid": 370, "correlation": 1932
    }
  },
  {
    "ph": "f", "id": 1932, "pid": 537, "tid": 537, "ts": 7429564851924.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851926.860, "dur": 4.232,
    "args": {
      "External id": 5252,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1933,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1933, "pid": 537, "tid": 537, "ts": 7429564851926.860,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851942.832, "dur": 5.657,
    "args": {
      "External id": 5255,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 1934,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1934, "pid": 537, "tid": 537, "ts": 7429564851942.832,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851962.052, "dur": 3.536,
    "args": {
      "External id": 5262,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1935,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1935, "pid": 537, "tid": 537, "ts": 7429564851962.052,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564851976.370, "dur": 5.318,
    "args": {
      "External id": 5265,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1936,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1936, "pid": 537, "tid": 537, "ts": 7429564851976.370,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852020.490, "dur": 0.236,
    "args": {
      "External id": 5277,      "cid": 370, "correlation": 1937
    }
  },
  {
    "ph": "f", "id": 1937, "pid": 537, "tid": 537, "ts": 7429564852020.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852023.942, "dur": 0.148,
    "args": {
      "External id": 5277,      "cid": 370, "correlation": 1938
    }
  },
  {
    "ph": "f", "id": 1938, "pid": 537, "tid": 537, "ts": 7429564852023.942,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852026.564, "dur": 0.188,
    "args": {
      "External id": 5277,      "cid": 370, "correlation": 1939
    }
  },
  {
    "ph": "f", "id": 1939, "pid": 537, "tid": 537, "ts": 7429564852026.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852028.654, "dur": 4.968,
    "args": {
      "External id": 5277,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1940,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1940, "pid": 537, "tid": 537, "ts": 7429564852028.654,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852049.620, "dur": 4.376,
    "args": {
      "External id": 5286,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1941,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1941, "pid": 537, "tid": 537, "ts": 7429564852049.620,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852095.566, "dur": 0.263,
    "args": {
      "External id": 5296,      "cid": 370, "correlation": 1942
    }
  },
  {
    "ph": "f", "id": 1942, "pid": 537, "tid": 537, "ts": 7429564852095.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852098.947, "dur": 0.159,
    "args": {
      "External id": 5296,      "cid": 370, "correlation": 1943
    }
  },
  {
    "ph": "f", "id": 1943, "pid": 537, "tid": 537, "ts": 7429564852098.947,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852101.229, "dur": 0.152,
    "args": {
      "External id": 5296,      "cid": 370, "correlation": 1944
    }
  },
  {
    "ph": "f", "id": 1944, "pid": 537, "tid": 537, "ts": 7429564852101.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852103.072, "dur": 4.338,
    "args": {
      "External id": 5296,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1945,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1945, "pid": 537, "tid": 537, "ts": 7429564852103.072,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852119.583, "dur": 5.599,
    "args": {
      "External id": 5299,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1946,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1946, "pid": 537, "tid": 537, "ts": 7429564852119.583,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852142.750, "dur": 4.189,
    "args": {
      "External id": 5301,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1947,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1947, "pid": 537, "tid": 537, "ts": 7429564852142.750,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852156.783, "dur": 4.439,
    "args": {
      "External id": 5304,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1948,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1948, "pid": 537, "tid": 537, "ts": 7429564852156.783,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852169.611, "dur": 4.750,
    "args": {
      "External id": 5305,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1949,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1949, "pid": 537, "tid": 537, "ts": 7429564852169.611,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852181.850, "dur": 5.300,
    "args": {
      "External id": 5306,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1950,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1950, "pid": 537, "tid": 537, "ts": 7429564852181.850,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852193.514, "dur": 3.828,
    "args": {
      "External id": 5307,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1951,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1951, "pid": 537, "tid": 537, "ts": 7429564852193.514,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852205.767, "dur": 3.492,
    "args": {
      "External id": 5309,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1952,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1952, "pid": 537, "tid": 537, "ts": 7429564852205.767,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852245.523, "dur": 0.247,
    "args": {
      "External id": 5317,      "cid": 370, "correlation": 1953
    }
  },
  {
    "ph": "f", "id": 1953, "pid": 537, "tid": 537, "ts": 7429564852245.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852248.837, "dur": 0.141,
    "args": {
      "External id": 5317,      "cid": 370, "correlation": 1954
    }
  },
  {
    "ph": "f", "id": 1954, "pid": 537, "tid": 537, "ts": 7429564852248.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852251.018, "dur": 0.144,
    "args": {
      "External id": 5317,      "cid": 370, "correlation": 1955
    }
  },
  {
    "ph": "f", "id": 1955, "pid": 537, "tid": 537, "ts": 7429564852251.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852257.953, "dur": 5.161,
    "args": {
      "External id": 5317,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1956,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1956, "pid": 537, "tid": 537, "ts": 7429564852257.953,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852263.464, "dur": 5.158,
    "args": {
      "External id": 5317,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1957,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1957, "pid": 537, "tid": 537, "ts": 7429564852263.464,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564852285.491, "dur": 0.329,
    "args": {
      "External id": 5320,      "cid": 290, "correlation": 1958
    }
  },
  {
    "ph": "f", "id": 1958, "pid": 537, "tid": 537, "ts": 7429564852285.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564852286.186, "dur": 47.554,
    "args": {
      "External id": 5320,            "bytes": 2097152, "cid": 108, "correlation": 1959, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 1959, "pid": 537, "tid": 537, "ts": 7429564852286.186,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852336.829, "dur": 5.479,
    "args": {
      "External id": 5320,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1960,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1960, "pid": 537, "tid": 537, "ts": 7429564852336.829,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852372.953, "dur": 0.220,
    "args": {
      "External id": 5328,      "cid": 370, "correlation": 1961
    }
  },
  {
    "ph": "f", "id": 1961, "pid": 537, "tid": 537, "ts": 7429564852372.953,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852375.925, "dur": 0.144,
    "args": {
      "External id": 5328,      "cid": 370, "correlation": 1962
    }
  },
  {
    "ph": "f", "id": 1962, "pid": 537, "tid": 537, "ts": 7429564852375.925,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852377.956, "dur": 0.156,
    "args": {
      "External id": 5328,      "cid": 370, "correlation": 1963
    }
  },
  {
    "ph": "f", "id": 1963, "pid": 537, "tid": 537, "ts": 7429564852377.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852382.927, "dur": 6.151,
    "args": {
      "External id": 5328,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1964,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1964, "pid": 537, "tid": 537, "ts": 7429564852382.927,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852389.405, "dur": 3.696,
    "args": {
      "External id": 5328,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1965,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1965, "pid": 537, "tid": 537, "ts": 7429564852389.405,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564852401.768, "dur": 0.235,
    "args": {
      "External id": 5331,      "cid": 290, "correlation": 1966
    }
  },
  {
    "ph": "f", "id": 1966, "pid": 537, "tid": 537, "ts": 7429564852401.768,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564852402.284, "dur": 30.399,
    "args": {
      "External id": 5331,            "bytes": 2097152, "cid": 108, "correlation": 1967, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 1967, "pid": 537, "tid": 537, "ts": 7429564852402.284,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852435.017, "dur": 4.190,
    "args": {
      "External id": 5331,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 1968,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1968, "pid": 537, "tid": 537, "ts": 7429564852435.017,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852465.900, "dur": 0.208,
    "args": {
      "External id": 5339,      "cid": 370, "correlation": 1969
    }
  },
  {
    "ph": "f", "id": 1969, "pid": 537, "tid": 537, "ts": 7429564852465.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852469.420, "dur": 0.149,
    "args": {
      "External id": 5339,      "cid": 370, "correlation": 1970
    }
  },
  {
    "ph": "f", "id": 1970, "pid": 537, "tid": 537, "ts": 7429564852469.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852471.686, "dur": 0.145,
    "args": {
      "External id": 5339,      "cid": 370, "correlation": 1971
    }
  },
  {
    "ph": "f", "id": 1971, "pid": 537, "tid": 537, "ts": 7429564852471.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852476.094, "dur": 5.776,
    "args": {
      "External id": 5339,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1972,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1972, "pid": 537, "tid": 537, "ts": 7429564852476.094,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852482.186, "dur": 4.256,
    "args": {
      "External id": 5339,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1973,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1973, "pid": 537, "tid": 537, "ts": 7429564852482.186,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852497.098, "dur": 4.704,
    "args": {
      "External id": 5342,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 1974,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1974, "pid": 537, "tid": 537, "ts": 7429564852497.098,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852531.011, "dur": 5.620,
    "args": {
      "External id": 5344,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1975,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1975, "pid": 537, "tid": 537, "ts": 7429564852531.011,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852546.284, "dur": 4.665,
    "args": {
      "External id": 5347,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 1976,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1976, "pid": 537, "tid": 537, "ts": 7429564852546.284,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852559.306, "dur": 3.976,
    "args": {
      "External id": 5348,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 1977,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1977, "pid": 537, "tid": 537, "ts": 7429564852559.306,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852569.516, "dur": 3.964,
    "args": {
      "External id": 5349,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 1978,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1978, "pid": 537, "tid": 537, "ts": 7429564852569.516,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852579.510, "dur": 3.345,
    "args": {
      "External id": 5350,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1979,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1979, "pid": 537, "tid": 537, "ts": 7429564852579.510,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852592.658, "dur": 4.944,
    "args": {
      "External id": 5352,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1980,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1980, "pid": 537, "tid": 537, "ts": 7429564852592.658,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852639.162, "dur": 0.192,
    "args": {
      "External id": 5360,      "cid": 370, "correlation": 1981
    }
  },
  {
    "ph": "f", "id": 1981, "pid": 537, "tid": 537, "ts": 7429564852639.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852642.573, "dur": 0.144,
    "args": {
      "External id": 5360,      "cid": 370, "correlation": 1982
    }
  },
  {
    "ph": "f", "id": 1982, "pid": 537, "tid": 537, "ts": 7429564852642.573,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852644.911, "dur": 0.170,
    "args": {
      "External id": 5360,      "cid": 370, "correlation": 1983
    }
  },
  {
    "ph": "f", "id": 1983, "pid": 537, "tid": 537, "ts": 7429564852644.911,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852647.082, "dur": 6.005,
    "args": {
      "External id": 5360,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1984,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1984, "pid": 537, "tid": 537, "ts": 7429564852647.082,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852689.636, "dur": 0.198,
    "args": {
      "External id": 5373,      "cid": 370, "correlation": 1985
    }
  },
  {
    "ph": "f", "id": 1985, "pid": 537, "tid": 537, "ts": 7429564852689.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852692.837, "dur": 0.142,
    "args": {
      "External id": 5373,      "cid": 370, "correlation": 1986
    }
  },
  {
    "ph": "f", "id": 1986, "pid": 537, "tid": 537, "ts": 7429564852692.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852694.930, "dur": 0.153,
    "args": {
      "External id": 5373,      "cid": 370, "correlation": 1987
    }
  },
  {
    "ph": "f", "id": 1987, "pid": 537, "tid": 537, "ts": 7429564852694.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852699.184, "dur": 5.632,
    "args": {
      "External id": 5373,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1988,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1988, "pid": 537, "tid": 537, "ts": 7429564852699.184,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852705.255, "dur": 3.935,
    "args": {
      "External id": 5373,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1989,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1989, "pid": 537, "tid": 537, "ts": 7429564852705.255,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852745.530, "dur": 0.164,
    "args": {
      "External id": 5386,      "cid": 370, "correlation": 1990
    }
  },
  {
    "ph": "f", "id": 1990, "pid": 537, "tid": 537, "ts": 7429564852745.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852748.051, "dur": 0.145,
    "args": {
      "External id": 5386,      "cid": 370, "correlation": 1991
    }
  },
  {
    "ph": "f", "id": 1991, "pid": 537, "tid": 537, "ts": 7429564852748.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564852750.104, "dur": 0.154,
    "args": {
      "External id": 5386,      "cid": 370, "correlation": 1992
    }
  },
  {
    "ph": "f", "id": 1992, "pid": 537, "tid": 537, "ts": 7429564852750.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852753.653, "dur": 4.187,
    "args": {
      "External id": 5386,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1993,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1993, "pid": 537, "tid": 537, "ts": 7429564852753.653,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852758.330, "dur": 4.828,
    "args": {
      "External id": 5386,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 1994,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1994, "pid": 537, "tid": 537, "ts": 7429564852758.330,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852785.612, "dur": 4.949,
    "args": {
      "External id": 5396,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1995,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1995, "pid": 537, "tid": 537, "ts": 7429564852785.612,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852813.182, "dur": 5.648,
    "args": {
      "External id": 5401,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 1996,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1996, "pid": 537, "tid": 537, "ts": 7429564852813.182,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852832.933, "dur": 6.115,
    "args": {
      "External id": 5402,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 1997,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1997, "pid": 537, "tid": 537, "ts": 7429564852832.933,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852846.709, "dur": 4.004,
    "args": {
      "External id": 5403,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 1998,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1998, "pid": 537, "tid": 537, "ts": 7429564852846.709,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852855.960, "dur": 4.110,
    "args": {
      "External id": 5404,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 1999,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1999, "pid": 537, "tid": 537, "ts": 7429564852855.960,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852866.685, "dur": 4.344,
    "args": {
      "External id": 5405,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2000,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2000, "pid": 537, "tid": 537, "ts": 7429564852866.685,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852883.310, "dur": 4.451,
    "args": {
      "External id": 5410,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2001,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2001, "pid": 537, "tid": 537, "ts": 7429564852883.310,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852894.871, "dur": 6.101,
    "args": {
      "External id": 5411,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2002,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2002, "pid": 537, "tid": 537, "ts": 7429564852894.871,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852906.899, "dur": 3.406,
    "args": {
      "External id": 5412,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2003,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2003, "pid": 537, "tid": 537, "ts": 7429564852906.899,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852915.370, "dur": 4.765,
    "args": {
      "External id": 5413,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2004,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2004, "pid": 537, "tid": 537, "ts": 7429564852915.370,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564852979.330, "dur": 4.756,
    "args": {
      "External id": 5430,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2005,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2005, "pid": 537, "tid": 537, "ts": 7429564852979.330,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853013.678, "dur": 4.207,
    "args": {
      "External id": 5448,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2006,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2006, "pid": 537, "tid": 537, "ts": 7429564853013.678,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853053.954, "dur": 0.324,
    "args": {
      "External id": 5461,      "cid": 370, "correlation": 2007
    }
  },
  {
    "ph": "f", "id": 2007, "pid": 537, "tid": 537, "ts": 7429564853053.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853058.004, "dur": 0.146,
    "args": {
      "External id": 5461,      "cid": 370, "correlation": 2008
    }
  },
  {
    "ph": "f", "id": 2008, "pid": 537, "tid": 537, "ts": 7429564853058.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853060.167, "dur": 0.186,
    "args": {
      "External id": 5461,      "cid": 370, "correlation": 2009
    }
  },
  {
    "ph": "f", "id": 2009, "pid": 537, "tid": 537, "ts": 7429564853060.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853062.742, "dur": 4.812,
    "args": {
      "External id": 5461,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2010,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2010, "pid": 537, "tid": 537, "ts": 7429564853062.742,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853079.358, "dur": 4.322,
    "args": {
      "External id": 5464,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 2011,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2011, "pid": 537, "tid": 537, "ts": 7429564853079.358,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853097.595, "dur": 3.753,
    "args": {
      "External id": 5471,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2012,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2012, "pid": 537, "tid": 537, "ts": 7429564853097.595,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853113.040, "dur": 5.583,
    "args": {
      "External id": 5474,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2013,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2013, "pid": 537, "tid": 537, "ts": 7429564853113.040,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853159.538, "dur": 0.207,
    "args": {
      "External id": 5486,      "cid": 370, "correlation": 2014
    }
  },
  {
    "ph": "f", "id": 2014, "pid": 537, "tid": 537, "ts": 7429564853159.538,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853162.990, "dur": 0.159,
    "args": {
      "External id": 5486,      "cid": 370, "correlation": 2015
    }
  },
  {
    "ph": "f", "id": 2015, "pid": 537, "tid": 537, "ts": 7429564853162.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853165.579, "dur": 0.162,
    "args": {
      "External id": 5486,      "cid": 370, "correlation": 2016
    }
  },
  {
    "ph": "f", "id": 2016, "pid": 537, "tid": 537, "ts": 7429564853165.579,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853167.602, "dur": 4.475,
    "args": {
      "External id": 5486,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2017,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2017, "pid": 537, "tid": 537, "ts": 7429564853167.602,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853188.383, "dur": 4.840,
    "args": {
      "External id": 5495,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2018,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2018, "pid": 537, "tid": 537, "ts": 7429564853188.383,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853233.750, "dur": 0.257,
    "args": {
      "External id": 5505,      "cid": 370, "correlation": 2019
    }
  },
  {
    "ph": "f", "id": 2019, "pid": 537, "tid": 537, "ts": 7429564853233.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853236.771, "dur": 0.143,
    "args": {
      "External id": 5505,      "cid": 370, "correlation": 2020
    }
  },
  {
    "ph": "f", "id": 2020, "pid": 537, "tid": 537, "ts": 7429564853236.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853239.024, "dur": 0.154,
    "args": {
      "External id": 5505,      "cid": 370, "correlation": 2021
    }
  },
  {
    "ph": "f", "id": 2021, "pid": 537, "tid": 537, "ts": 7429564853239.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853240.913, "dur": 6.232,
    "args": {
      "External id": 5505,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2022,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2022, "pid": 537, "tid": 537, "ts": 7429564853240.913,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564853259.027, "dur": 0.380,
    "args": {
      "External id": 5508,      "cid": 290, "correlation": 2023
    }
  },
  {
    "ph": "f", "id": 2023, "pid": 537, "tid": 537, "ts": 7429564853259.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564853259.853, "dur": 52.524,
    "args": {
      "External id": 5508,            "bytes": 2097152, "cid": 108, "correlation": 2024, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 2024, "pid": 537, "tid": 537, "ts": 7429564853259.853,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853314.959, "dur": 5.575,
    "args": {
      "External id": 5508,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2025,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2025, "pid": 537, "tid": 537, "ts": 7429564853314.959,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853342.160, "dur": 5.497,
    "args": {
      "External id": 5510,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2026,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2026, "pid": 537, "tid": 537, "ts": 7429564853342.160,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853357.776, "dur": 6.571,
    "args": {
      "External id": 5513,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2027,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2027, "pid": 537, "tid": 537, "ts": 7429564853357.776,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853373.848, "dur": 3.811,
    "args": {
      "External id": 5514,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2028,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2028, "pid": 537, "tid": 537, "ts": 7429564853373.848,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853383.577, "dur": 4.939,
    "args": {
      "External id": 5515,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2029,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2029, "pid": 537, "tid": 537, "ts": 7429564853383.577,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853396.375, "dur": 5.476,
    "args": {
      "External id": 5516,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2030,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2030, "pid": 537, "tid": 537, "ts": 7429564853396.375,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853409.456, "dur": 5.276,
    "args": {
      "External id": 5518,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2031,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2031, "pid": 537, "tid": 537, "ts": 7429564853409.456,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853453.706, "dur": 0.248,
    "args": {
      "External id": 5526,      "cid": 370, "correlation": 2032
    }
  },
  {
    "ph": "f", "id": 2032, "pid": 537, "tid": 537, "ts": 7429564853453.706,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853457.853, "dur": 0.144,
    "args": {
      "External id": 5526,      "cid": 370, "correlation": 2033
    }
  },
  {
    "ph": "f", "id": 2033, "pid": 537, "tid": 537, "ts": 7429564853457.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853460.594, "dur": 0.140,
    "args": {
      "External id": 5526,      "cid": 370, "correlation": 2034
    }
  },
  {
    "ph": "f", "id": 2034, "pid": 537, "tid": 537, "ts": 7429564853460.594,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853468.228, "dur": 5.545,
    "args": {
      "External id": 5526,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2035,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2035, "pid": 537, "tid": 537, "ts": 7429564853468.228,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853474.326, "dur": 3.639,
    "args": {
      "External id": 5526,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2036,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2036, "pid": 537, "tid": 537, "ts": 7429564853474.326,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853495.309, "dur": 4.364,
    "args": {
      "External id": 5529,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2037,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2037, "pid": 537, "tid": 537, "ts": 7429564853495.309,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564853519.545, "dur": 0.298,
    "args": {
      "External id": 5537,      "cid": 290, "correlation": 2038
    }
  },
  {
    "ph": "f", "id": 2038, "pid": 537, "tid": 537, "ts": 7429564853519.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564853520.191, "dur": 36.275,
    "args": {
      "External id": 5537,            "bytes": 2097152, "cid": 108, "correlation": 2039, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 2039, "pid": 537, "tid": 537, "ts": 7429564853520.191,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853569.661, "dur": 0.186,
    "args": {
      "External id": 5537,      "cid": 370, "correlation": 2040
    }
  },
  {
    "ph": "f", "id": 2040, "pid": 537, "tid": 537, "ts": 7429564853569.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853572.758, "dur": 0.144,
    "args": {
      "External id": 5537,      "cid": 370, "correlation": 2041
    }
  },
  {
    "ph": "f", "id": 2041, "pid": 537, "tid": 537, "ts": 7429564853572.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853575.345, "dur": 0.147,
    "args": {
      "External id": 5537,      "cid": 370, "correlation": 2042
    }
  },
  {
    "ph": "f", "id": 2042, "pid": 537, "tid": 537, "ts": 7429564853575.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853580.029, "dur": 5.992,
    "args": {
      "External id": 5537,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2043,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2043, "pid": 537, "tid": 537, "ts": 7429564853580.029,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853586.336, "dur": 4.459,
    "args": {
      "External id": 5537,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2044,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2044, "pid": 537, "tid": 537, "ts": 7429564853586.336,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853600.600, "dur": 5.174,
    "args": {
      "External id": 5540,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 2045,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2045, "pid": 537, "tid": 537, "ts": 7429564853600.600,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853633.623, "dur": 0.185,
    "args": {
      "External id": 5548,      "cid": 370, "correlation": 2046
    }
  },
  {
    "ph": "f", "id": 2046, "pid": 537, "tid": 537, "ts": 7429564853633.623,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853636.837, "dur": 0.144,
    "args": {
      "External id": 5548,      "cid": 370, "correlation": 2047
    }
  },
  {
    "ph": "f", "id": 2047, "pid": 537, "tid": 537, "ts": 7429564853636.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853638.783, "dur": 0.160,
    "args": {
      "External id": 5548,      "cid": 370, "correlation": 2048
    }
  },
  {
    "ph": "f", "id": 2048, "pid": 537, "tid": 537, "ts": 7429564853638.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853642.521, "dur": 5.307,
    "args": {
      "External id": 5548,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2049,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2049, "pid": 537, "tid": 537, "ts": 7429564853642.521,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853658.307, "dur": 4.258,
    "args": {
      "External id": 5548,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2050,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2050, "pid": 537, "tid": 537, "ts": 7429564853658.307,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853672.460, "dur": 4.155,
    "args": {
      "External id": 5551,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2051,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2051, "pid": 537, "tid": 537, "ts": 7429564853672.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853703.798, "dur": 3.806,
    "args": {
      "External id": 5553,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2052,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2052, "pid": 537, "tid": 537, "ts": 7429564853703.798,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853716.239, "dur": 4.187,
    "args": {
      "External id": 5556,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2053,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2053, "pid": 537, "tid": 537, "ts": 7429564853716.239,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853727.872, "dur": 4.977,
    "args": {
      "External id": 5557,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2054,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2054, "pid": 537, "tid": 537, "ts": 7429564853727.872,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853737.702, "dur": 4.226,
    "args": {
      "External id": 5558,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2055,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2055, "pid": 537, "tid": 537, "ts": 7429564853737.702,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853747.826, "dur": 3.430,
    "args": {
      "External id": 5559,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2056,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2056, "pid": 537, "tid": 537, "ts": 7429564853747.826,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564853758.424, "dur": 0.229,
    "args": {
      "External id": 5561,      "cid": 290, "correlation": 2057
    }
  },
  {
    "ph": "f", "id": 2057, "pid": 537, "tid": 537, "ts": 7429564853758.424,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564853759.097, "dur": 32.709,
    "args": {
      "External id": 5561,            "bytes": 2097152, "cid": 108, "correlation": 2058, "ptr": "0x100000002"
    }
  },
  {
    "ph": "s", "id": 2058, "pid": 537, "tid": 537, "ts": 7429564853759.097,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853793.691, "dur": 3.432,
    "args": {
      "External id": 5561,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2059,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2059, "pid": 537, "tid": 537, "ts": 7429564853793.691,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853836.056, "dur": 0.382,
    "args": {
      "External id": 5569,      "cid": 370, "correlation": 2060
    }
  },
  {
    "ph": "f", "id": 2060, "pid": 537, "tid": 537, "ts": 7429564853836.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853839.442, "dur": 0.144,
    "args": {
      "External id": 5569,      "cid": 370, "correlation": 2061
    }
  },
  {
    "ph": "f", "id": 2061, "pid": 537, "tid": 537, "ts": 7429564853839.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853841.570, "dur": 0.184,
    "args": {
      "External id": 5569,      "cid": 370, "correlation": 2062
    }
  },
  {
    "ph": "f", "id": 2062, "pid": 537, "tid": 537, "ts": 7429564853841.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853844.123, "dur": 4.902,
    "args": {
      "External id": 5569,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2063,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2063, "pid": 537, "tid": 537, "ts": 7429564853844.123,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853883.369, "dur": 0.185,
    "args": {
      "External id": 5582,      "cid": 370, "correlation": 2064
    }
  },
  {
    "ph": "f", "id": 2064, "pid": 537, "tid": 537, "ts": 7429564853883.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853886.124, "dur": 0.145,
    "args": {
      "External id": 5582,      "cid": 370, "correlation": 2065
    }
  },
  {
    "ph": "f", "id": 2065, "pid": 537, "tid": 537, "ts": 7429564853886.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853888.017, "dur": 0.175,
    "args": {
      "External id": 5582,      "cid": 370, "correlation": 2066
    }
  },
  {
    "ph": "f", "id": 2066, "pid": 537, "tid": 537, "ts": 7429564853888.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853892.566, "dur": 6.397,
    "args": {
      "External id": 5582,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2067,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2067, "pid": 537, "tid": 537, "ts": 7429564853892.566,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853899.379, "dur": 4.782,
    "args": {
      "External id": 5582,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2068,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2068, "pid": 537, "tid": 537, "ts": 7429564853899.379,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853935.343, "dur": 0.186,
    "args": {
      "External id": 5595,      "cid": 370, "correlation": 2069
    }
  },
  {
    "ph": "f", "id": 2069, "pid": 537, "tid": 537, "ts": 7429564853935.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853937.674, "dur": 0.145,
    "args": {
      "External id": 5595,      "cid": 370, "correlation": 2070
    }
  },
  {
    "ph": "f", "id": 2070, "pid": 537, "tid": 537, "ts": 7429564853937.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564853939.445, "dur": 0.177,
    "args": {
      "External id": 5595,      "cid": 370, "correlation": 2071
    }
  },
  {
    "ph": "f", "id": 2071, "pid": 537, "tid": 537, "ts": 7429564853939.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853942.922, "dur": 4.655,
    "args": {
      "External id": 5595,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2072,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2072, "pid": 537, "tid": 537, "ts": 7429564853942.922,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853948.030, "dur": 4.300,
    "args": {
      "External id": 5595,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2073,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2073, "pid": 537, "tid": 537, "ts": 7429564853948.030,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853975.536, "dur": 3.587,
    "args": {
      "External id": 5605,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2074,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2074, "pid": 537, "tid": 537, "ts": 7429564853975.536,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564853999.610, "dur": 4.455,
    "args": {
      "External id": 5610,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2075,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2075, "pid": 537, "tid": 537, "ts": 7429564853999.610,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854017.502, "dur": 6.451,
    "args": {
      "External id": 5611,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2076,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2076, "pid": 537, "tid": 537, "ts": 7429564854017.502,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854032.530, "dur": 4.912,
    "args": {
      "External id": 5612,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2077,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2077, "pid": 537, "tid": 537, "ts": 7429564854032.530,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854042.681, "dur": 4.066,
    "args": {
      "External id": 5613,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2078,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2078, "pid": 537, "tid": 537, "ts": 7429564854042.681,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854052.888, "dur": 4.141,
    "args": {
      "External id": 5614,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2079,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2079, "pid": 537, "tid": 537, "ts": 7429564854052.888,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854088.202, "dur": 4.310,
    "args": {
      "External id": 5619,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2080,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2080, "pid": 537, "tid": 537, "ts": 7429564854088.202,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854108.223, "dur": 5.995,
    "args": {
      "External id": 5620,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2081,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2081, "pid": 537, "tid": 537, "ts": 7429564854108.223,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854123.471, "dur": 4.083,
    "args": {
      "External id": 5621,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2082,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2082, "pid": 537, "tid": 537, "ts": 7429564854123.471,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854133.303, "dur": 5.367,
    "args": {
      "External id": 5622,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2083,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2083, "pid": 537, "tid": 537, "ts": 7429564854133.303,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854228.157, "dur": 5.098,
    "args": {
      "External id": 5639,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2084,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2084, "pid": 537, "tid": 537, "ts": 7429564854228.157,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854261.381, "dur": 4.032,
    "args": {
      "External id": 5657,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2085,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2085, "pid": 537, "tid": 537, "ts": 7429564854261.381,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854304.726, "dur": 0.399,
    "args": {
      "External id": 5670,      "cid": 370, "correlation": 2086
    }
  },
  {
    "ph": "f", "id": 2086, "pid": 537, "tid": 537, "ts": 7429564854304.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854310.298, "dur": 0.149,
    "args": {
      "External id": 5670,      "cid": 370, "correlation": 2087
    }
  },
  {
    "ph": "f", "id": 2087, "pid": 537, "tid": 537, "ts": 7429564854310.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854313.248, "dur": 0.149,
    "args": {
      "External id": 5670,      "cid": 370, "correlation": 2088
    }
  },
  {
    "ph": "f", "id": 2088, "pid": 537, "tid": 537, "ts": 7429564854313.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854317.539, "dur": 4.848,
    "args": {
      "External id": 5670,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2089,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2089, "pid": 537, "tid": 537, "ts": 7429564854317.539,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854336.527, "dur": 4.600,
    "args": {
      "External id": 5673,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 2090,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2090, "pid": 537, "tid": 537, "ts": 7429564854336.527,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854355.434, "dur": 4.540,
    "args": {
      "External id": 5680,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2091,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2091, "pid": 537, "tid": 537, "ts": 7429564854355.434,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854372.294, "dur": 7.002,
    "args": {
      "External id": 5683,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2092,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2092, "pid": 537, "tid": 537, "ts": 7429564854372.294,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854421.136, "dur": 0.216,
    "args": {
      "External id": 5695,      "cid": 370, "correlation": 2093
    }
  },
  {
    "ph": "f", "id": 2093, "pid": 537, "tid": 537, "ts": 7429564854421.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854424.729, "dur": 0.155,
    "args": {
      "External id": 5695,      "cid": 370, "correlation": 2094
    }
  },
  {
    "ph": "f", "id": 2094, "pid": 537, "tid": 537, "ts": 7429564854424.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854427.674, "dur": 0.165,
    "args": {
      "External id": 5695,      "cid": 370, "correlation": 2095
    }
  },
  {
    "ph": "f", "id": 2095, "pid": 537, "tid": 537, "ts": 7429564854427.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854430.123, "dur": 5.217,
    "args": {
      "External id": 5695,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2096,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2096, "pid": 537, "tid": 537, "ts": 7429564854430.123,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854451.994, "dur": 4.395,
    "args": {
      "External id": 5704,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2097,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2097, "pid": 537, "tid": 537, "ts": 7429564854451.994,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854501.522, "dur": 0.264,
    "args": {
      "External id": 5714,      "cid": 370, "correlation": 2098
    }
  },
  {
    "ph": "f", "id": 2098, "pid": 537, "tid": 537, "ts": 7429564854501.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854504.868, "dur": 0.179,
    "args": {
      "External id": 5714,      "cid": 370, "correlation": 2099
    }
  },
  {
    "ph": "f", "id": 2099, "pid": 537, "tid": 537, "ts": 7429564854504.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854507.077, "dur": 0.194,
    "args": {
      "External id": 5714,      "cid": 370, "correlation": 2100
    }
  },
  {
    "ph": "f", "id": 2100, "pid": 537, "tid": 537, "ts": 7429564854507.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854509.077, "dur": 4.233,
    "args": {
      "External id": 5714,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2101,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2101, "pid": 537, "tid": 537, "ts": 7429564854509.077,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854526.072, "dur": 4.636,
    "args": {
      "External id": 5717,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2102,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2102, "pid": 537, "tid": 537, "ts": 7429564854526.072,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854549.362, "dur": 5.719,
    "args": {
      "External id": 5719,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2103,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2103, "pid": 537, "tid": 537, "ts": 7429564854549.362,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854566.050, "dur": 5.552,
    "args": {
      "External id": 5722,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2104,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2104, "pid": 537, "tid": 537, "ts": 7429564854566.050,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854580.482, "dur": 4.127,
    "args": {
      "External id": 5723,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2105,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2105, "pid": 537, "tid": 537, "ts": 7429564854580.482,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854592.104, "dur": 4.172,
    "args": {
      "External id": 5724,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2106,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2106, "pid": 537, "tid": 537, "ts": 7429564854592.104,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854602.733, "dur": 4.228,
    "args": {
      "External id": 5725,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2107,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2107, "pid": 537, "tid": 537, "ts": 7429564854602.733,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854615.093, "dur": 4.230,
    "args": {
      "External id": 5727,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2108,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2108, "pid": 537, "tid": 537, "ts": 7429564854615.093,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564854645.993, "dur": 0.461,
    "args": {
      "External id": 5735,      "cid": 290, "correlation": 2109
    }
  },
  {
    "ph": "f", "id": 2109, "pid": 537, "tid": 537, "ts": 7429564854645.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564854646.986, "dur": 60.180,
    "args": {
      "External id": 5735,            "bytes": 2097152, "cid": 108, "correlation": 2110, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 2110, "pid": 537, "tid": 537, "ts": 7429564854646.986,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854720.329, "dur": 0.291,
    "args": {
      "External id": 5735,      "cid": 370, "correlation": 2111
    }
  },
  {
    "ph": "f", "id": 2111, "pid": 537, "tid": 537, "ts": 7429564854720.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854724.529, "dur": 0.166,
    "args": {
      "External id": 5735,      "cid": 370, "correlation": 2112
    }
  },
  {
    "ph": "f", "id": 2112, "pid": 537, "tid": 537, "ts": 7429564854724.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854726.973, "dur": 0.167,
    "args": {
      "External id": 5735,      "cid": 370, "correlation": 2113
    }
  },
  {
    "ph": "f", "id": 2113, "pid": 537, "tid": 537, "ts": 7429564854726.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854736.639, "dur": 6.159,
    "args": {
      "External id": 5735,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2114,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2114, "pid": 537, "tid": 537, "ts": 7429564854736.639,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854743.271, "dur": 4.685,
    "args": {
      "External id": 5735,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2115,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2115, "pid": 537, "tid": 537, "ts": 7429564854743.271,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854768.751, "dur": 5.362,
    "args": {
      "External id": 5738,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2116,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2116, "pid": 537, "tid": 537, "ts": 7429564854768.751,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564854793.074, "dur": 0.306,
    "args": {
      "External id": 5746,      "cid": 290, "correlation": 2117
    }
  },
  {
    "ph": "f", "id": 2117, "pid": 537, "tid": 537, "ts": 7429564854793.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564854793.701, "dur": 32.084,
    "args": {
      "External id": 5746,            "bytes": 2097152, "cid": 108, "correlation": 2118, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 2118, "pid": 537, "tid": 537, "ts": 7429564854793.701,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854836.078, "dur": 0.212,
    "args": {
      "External id": 5746,      "cid": 370, "correlation": 2119
    }
  },
  {
    "ph": "f", "id": 2119, "pid": 537, "tid": 537, "ts": 7429564854836.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854839.061, "dur": 0.140,
    "args": {
      "External id": 5746,      "cid": 370, "correlation": 2120
    }
  },
  {
    "ph": "f", "id": 2120, "pid": 537, "tid": 537, "ts": 7429564854839.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854841.360, "dur": 0.151,
    "args": {
      "External id": 5746,      "cid": 370, "correlation": 2121
    }
  },
  {
    "ph": "f", "id": 2121, "pid": 537, "tid": 537, "ts": 7429564854841.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854846.105, "dur": 5.943,
    "args": {
      "External id": 5746,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2122,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2122, "pid": 537, "tid": 537, "ts": 7429564854846.105,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854852.406, "dur": 3.598,
    "args": {
      "External id": 5746,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2123,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2123, "pid": 537, "tid": 537, "ts": 7429564854852.406,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854866.228, "dur": 4.255,
    "args": {
      "External id": 5749,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 2124,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2124, "pid": 537, "tid": 537, "ts": 7429564854866.228,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854896.640, "dur": 0.172,
    "args": {
      "External id": 5757,      "cid": 370, "correlation": 2125
    }
  },
  {
    "ph": "f", "id": 2125, "pid": 537, "tid": 537, "ts": 7429564854896.640,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854899.598, "dur": 0.142,
    "args": {
      "External id": 5757,      "cid": 370, "correlation": 2126
    }
  },
  {
    "ph": "f", "id": 2126, "pid": 537, "tid": 537, "ts": 7429564854899.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564854901.377, "dur": 0.150,
    "args": {
      "External id": 5757,      "cid": 370, "correlation": 2127
    }
  },
  {
    "ph": "f", "id": 2127, "pid": 537, "tid": 537, "ts": 7429564854901.377,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854905.118, "dur": 4.935,
    "args": {
      "External id": 5757,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2128,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2128, "pid": 537, "tid": 537, "ts": 7429564854905.118,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854910.593, "dur": 4.538,
    "args": {
      "External id": 5757,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2129,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2129, "pid": 537, "tid": 537, "ts": 7429564854910.593,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854925.397, "dur": 4.591,
    "args": {
      "External id": 5760,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2130,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2130, "pid": 537, "tid": 537, "ts": 7429564854925.397,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854959.519, "dur": 4.440,
    "args": {
      "External id": 5762,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2131,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2131, "pid": 537, "tid": 537, "ts": 7429564854959.519,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854973.826, "dur": 4.924,
    "args": {
      "External id": 5765,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2132,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2132, "pid": 537, "tid": 537, "ts": 7429564854973.826,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854988.322, "dur": 5.158,
    "args": {
      "External id": 5766,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2133,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2133, "pid": 537, "tid": 537, "ts": 7429564854988.322,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564854998.821, "dur": 4.606,
    "args": {
      "External id": 5767,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2134,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2134, "pid": 537, "tid": 537, "ts": 7429564854998.821,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855009.798, "dur": 6.316,
    "args": {
      "External id": 5768,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2135,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2135, "pid": 537, "tid": 537, "ts": 7429564855009.798,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855024.265, "dur": 3.679,
    "args": {
      "External id": 5770,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2136,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2136, "pid": 537, "tid": 537, "ts": 7429564855024.265,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855067.382, "dur": 0.256,
    "args": {
      "External id": 5778,      "cid": 370, "correlation": 2137
    }
  },
  {
    "ph": "f", "id": 2137, "pid": 537, "tid": 537, "ts": 7429564855067.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855070.856, "dur": 0.175,
    "args": {
      "External id": 5778,      "cid": 370, "correlation": 2138
    }
  },
  {
    "ph": "f", "id": 2138, "pid": 537, "tid": 537, "ts": 7429564855070.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855073.061, "dur": 0.168,
    "args": {
      "External id": 5778,      "cid": 370, "correlation": 2139
    }
  },
  {
    "ph": "f", "id": 2139, "pid": 537, "tid": 537, "ts": 7429564855073.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855075.092, "dur": 4.943,
    "args": {
      "External id": 5778,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2140,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2140, "pid": 537, "tid": 537, "ts": 7429564855075.092,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855117.386, "dur": 0.162,
    "args": {
      "External id": 5791,      "cid": 370, "correlation": 2141
    }
  },
  {
    "ph": "f", "id": 2141, "pid": 537, "tid": 537, "ts": 7429564855117.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855120.550, "dur": 0.155,
    "args": {
      "External id": 5791,      "cid": 370, "correlation": 2142
    }
  },
  {
    "ph": "f", "id": 2142, "pid": 537, "tid": 537, "ts": 7429564855120.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855122.688, "dur": 0.156,
    "args": {
      "External id": 5791,      "cid": 370, "correlation": 2143
    }
  },
  {
    "ph": "f", "id": 2143, "pid": 537, "tid": 537, "ts": 7429564855122.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855127.036, "dur": 5.740,
    "args": {
      "External id": 5791,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2144,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2144, "pid": 537, "tid": 537, "ts": 7429564855127.036,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855133.288, "dur": 3.626,
    "args": {
      "External id": 5791,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2145,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2145, "pid": 537, "tid": 537, "ts": 7429564855133.288,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855172.664, "dur": 0.279,
    "args": {
      "External id": 5804,      "cid": 370, "correlation": 2146
    }
  },
  {
    "ph": "f", "id": 2146, "pid": 537, "tid": 537, "ts": 7429564855172.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855175.379, "dur": 0.142,
    "args": {
      "External id": 5804,      "cid": 370, "correlation": 2147
    }
  },
  {
    "ph": "f", "id": 2147, "pid": 537, "tid": 537, "ts": 7429564855175.379,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855177.187, "dur": 0.156,
    "args": {
      "External id": 5804,      "cid": 370, "correlation": 2148
    }
  },
  {
    "ph": "f", "id": 2148, "pid": 537, "tid": 537, "ts": 7429564855177.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855180.555, "dur": 4.912,
    "args": {
      "External id": 5804,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2149,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2149, "pid": 537, "tid": 537, "ts": 7429564855180.555,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855185.761, "dur": 4.561,
    "args": {
      "External id": 5804,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2150,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2150, "pid": 537, "tid": 537, "ts": 7429564855185.761,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855213.287, "dur": 3.875,
    "args": {
      "External id": 5814,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2151,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2151, "pid": 537, "tid": 537, "ts": 7429564855213.287,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855239.573, "dur": 4.245,
    "args": {
      "External id": 5819,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2152,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2152, "pid": 537, "tid": 537, "ts": 7429564855239.573,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564855255.456, "dur": 0.373,
    "args": {
      "External id": 5820,      "cid": 290, "correlation": 2153
    }
  },
  {
    "ph": "f", "id": 2153, "pid": 537, "tid": 537, "ts": 7429564855255.456,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564855256.373, "dur": 37.892,
    "args": {
      "External id": 5820,            "bytes": 2097152, "cid": 108, "correlation": 2154, "ptr": "0x5569ba9ecb70"
    }
  },
  {
    "ph": "s", "id": 2154, "pid": 537, "tid": 537, "ts": 7429564855256.373,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855298.513, "dur": 6.418,
    "args": {
      "External id": 5820,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2155,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2155, "pid": 537, "tid": 537, "ts": 7429564855298.513,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855313.287, "dur": 3.916,
    "args": {
      "External id": 5821,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2156,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2156, "pid": 537, "tid": 537, "ts": 7429564855313.287,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855323.853, "dur": 4.253,
    "args": {
      "External id": 5822,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2157,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2157, "pid": 537, "tid": 537, "ts": 7429564855323.853,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855333.853, "dur": 4.707,
    "args": {
      "External id": 5823,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2158,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2158, "pid": 537, "tid": 537, "ts": 7429564855333.853,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855353.849, "dur": 3.726,
    "args": {
      "External id": 5828,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2159,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2159, "pid": 537, "tid": 537, "ts": 7429564855353.849,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855364.675, "dur": 4.756,
    "args": {
      "External id": 5829,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2160,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2160, "pid": 537, "tid": 537, "ts": 7429564855364.675,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855376.430, "dur": 3.925,
    "args": {
      "External id": 5830,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2161,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2161, "pid": 537, "tid": 537, "ts": 7429564855376.430,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855385.097, "dur": 3.250,
    "args": {
      "External id": 5831,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2162,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2162, "pid": 537, "tid": 537, "ts": 7429564855385.097,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855446.559, "dur": 4.618,
    "args": {
      "External id": 5848,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2163,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2163, "pid": 537, "tid": 537, "ts": 7429564855446.559,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855480.861, "dur": 4.665,
    "args": {
      "External id": 5866,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2164,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2164, "pid": 537, "tid": 537, "ts": 7429564855480.861,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855521.576, "dur": 0.356,
    "args": {
      "External id": 5879,      "cid": 370, "correlation": 2165
    }
  },
  {
    "ph": "f", "id": 2165, "pid": 537, "tid": 537, "ts": 7429564855521.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855526.077, "dur": 0.167,
    "args": {
      "External id": 5879,      "cid": 370, "correlation": 2166
    }
  },
  {
    "ph": "f", "id": 2166, "pid": 537, "tid": 537, "ts": 7429564855526.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855528.792, "dur": 0.160,
    "args": {
      "External id": 5879,      "cid": 370, "correlation": 2167
    }
  },
  {
    "ph": "f", "id": 2167, "pid": 537, "tid": 537, "ts": 7429564855528.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855531.868, "dur": 5.414,
    "args": {
      "External id": 5879,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2168,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2168, "pid": 537, "tid": 537, "ts": 7429564855531.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855549.203, "dur": 4.624,
    "args": {
      "External id": 5882,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 2169,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2169, "pid": 537, "tid": 537, "ts": 7429564855549.203,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855568.058, "dur": 3.951,
    "args": {
      "External id": 5889,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2170,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2170, "pid": 537, "tid": 537, "ts": 7429564855568.058,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855583.429, "dur": 5.521,
    "args": {
      "External id": 5892,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2171,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2171, "pid": 537, "tid": 537, "ts": 7429564855583.429,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855633.088, "dur": 0.201,
    "args": {
      "External id": 5904,      "cid": 370, "correlation": 2172
    }
  },
  {
    "ph": "f", "id": 2172, "pid": 537, "tid": 537, "ts": 7429564855633.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855636.591, "dur": 0.149,
    "args": {
      "External id": 5904,      "cid": 370, "correlation": 2173
    }
  },
  {
    "ph": "f", "id": 2173, "pid": 537, "tid": 537, "ts": 7429564855636.591,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855639.524, "dur": 0.151,
    "args": {
      "External id": 5904,      "cid": 370, "correlation": 2174
    }
  },
  {
    "ph": "f", "id": 2174, "pid": 537, "tid": 537, "ts": 7429564855639.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855641.525, "dur": 4.762,
    "args": {
      "External id": 5904,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2175,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2175, "pid": 537, "tid": 537, "ts": 7429564855641.525,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855661.930, "dur": 5.046,
    "args": {
      "External id": 5913,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2176,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2176, "pid": 537, "tid": 537, "ts": 7429564855661.930,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855708.179, "dur": 0.272,
    "args": {
      "External id": 5923,      "cid": 370, "correlation": 2177
    }
  },
  {
    "ph": "f", "id": 2177, "pid": 537, "tid": 537, "ts": 7429564855708.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855711.618, "dur": 0.153,
    "args": {
      "External id": 5923,      "cid": 370, "correlation": 2178
    }
  },
  {
    "ph": "f", "id": 2178, "pid": 537, "tid": 537, "ts": 7429564855711.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855713.614, "dur": 0.194,
    "args": {
      "External id": 5923,      "cid": 370, "correlation": 2179
    }
  },
  {
    "ph": "f", "id": 2179, "pid": 537, "tid": 537, "ts": 7429564855713.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855715.468, "dur": 5.585,
    "args": {
      "External id": 5923,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2180,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2180, "pid": 537, "tid": 537, "ts": 7429564855715.468,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855733.922, "dur": 5.707,
    "args": {
      "External id": 5926,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2181,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2181, "pid": 537, "tid": 537, "ts": 7429564855733.922,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855757.095, "dur": 4.802,
    "args": {
      "External id": 5928,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2182,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2182, "pid": 537, "tid": 537, "ts": 7429564855757.095,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855771.643, "dur": 5.532,
    "args": {
      "External id": 5931,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2183,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2183, "pid": 537, "tid": 537, "ts": 7429564855771.643,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855785.025, "dur": 3.983,
    "args": {
      "External id": 5932,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2184,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2184, "pid": 537, "tid": 537, "ts": 7429564855785.025,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855794.934, "dur": 5.384,
    "args": {
      "External id": 5933,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2185,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2185, "pid": 537, "tid": 537, "ts": 7429564855794.934,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855808.449, "dur": 5.229,
    "args": {
      "External id": 5934,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2186,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2186, "pid": 537, "tid": 537, "ts": 7429564855808.449,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855821.252, "dur": 4.968,
    "args": {
      "External id": 5936,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2187,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2187, "pid": 537, "tid": 537, "ts": 7429564855821.252,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855863.175, "dur": 0.200,
    "args": {
      "External id": 5944,      "cid": 370, "correlation": 2188
    }
  },
  {
    "ph": "f", "id": 2188, "pid": 537, "tid": 537, "ts": 7429564855863.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855866.569, "dur": 0.145,
    "args": {
      "External id": 5944,      "cid": 370, "correlation": 2189
    }
  },
  {
    "ph": "f", "id": 2189, "pid": 537, "tid": 537, "ts": 7429564855866.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855868.954, "dur": 0.143,
    "args": {
      "External id": 5944,      "cid": 370, "correlation": 2190
    }
  },
  {
    "ph": "f", "id": 2190, "pid": 537, "tid": 537, "ts": 7429564855868.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855876.367, "dur": 4.534,
    "args": {
      "External id": 5944,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2191,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2191, "pid": 537, "tid": 537, "ts": 7429564855876.367,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855881.298, "dur": 4.710,
    "args": {
      "External id": 5944,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2192,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2192, "pid": 537, "tid": 537, "ts": 7429564855881.298,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564855902.709, "dur": 0.429,
    "args": {
      "External id": 5947,      "cid": 290, "correlation": 2193
    }
  },
  {
    "ph": "f", "id": 2193, "pid": 537, "tid": 537, "ts": 7429564855902.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564855903.597, "dur": 45.760,
    "args": {
      "External id": 5947,            "bytes": 2097152, "cid": 108, "correlation": 2194, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 2194, "pid": 537, "tid": 537, "ts": 7429564855903.597,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564855952.658, "dur": 6.300,
    "args": {
      "External id": 5947,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2195,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2195, "pid": 537, "tid": 537, "ts": 7429564855952.658,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855990.211, "dur": 0.221,
    "args": {
      "External id": 5955,      "cid": 370, "correlation": 2196
    }
  },
  {
    "ph": "f", "id": 2196, "pid": 537, "tid": 537, "ts": 7429564855990.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855994.310, "dur": 0.148,
    "args": {
      "External id": 5955,      "cid": 370, "correlation": 2197
    }
  },
  {
    "ph": "f", "id": 2197, "pid": 537, "tid": 537, "ts": 7429564855994.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564855996.849, "dur": 0.145,
    "args": {
      "External id": 5955,      "cid": 370, "correlation": 2198
    }
  },
  {
    "ph": "f", "id": 2198, "pid": 537, "tid": 537, "ts": 7429564855996.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856002.209, "dur": 6.097,
    "args": {
      "External id": 5955,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2199,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2199, "pid": 537, "tid": 537, "ts": 7429564856002.209,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856008.699, "dur": 4.027,
    "args": {
      "External id": 5955,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2200,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2200, "pid": 537, "tid": 537, "ts": 7429564856008.699,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564856022.164, "dur": 0.321,
    "args": {
      "External id": 5958,      "cid": 290, "correlation": 2201
    }
  },
  {
    "ph": "f", "id": 2201, "pid": 537, "tid": 537, "ts": 7429564856022.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564856022.927, "dur": 31.306,
    "args": {
      "External id": 5958,            "bytes": 2097152, "cid": 108, "correlation": 2202, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 2202, "pid": 537, "tid": 537, "ts": 7429564856022.927,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856056.459, "dur": 3.978,
    "args": {
      "External id": 5958,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 2203,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2203, "pid": 537, "tid": 537, "ts": 7429564856056.459,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856090.932, "dur": 0.188,
    "args": {
      "External id": 5966,      "cid": 370, "correlation": 2204
    }
  },
  {
    "ph": "f", "id": 2204, "pid": 537, "tid": 537, "ts": 7429564856090.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856094.734, "dur": 0.146,
    "args": {
      "External id": 5966,      "cid": 370, "correlation": 2205
    }
  },
  {
    "ph": "f", "id": 2205, "pid": 537, "tid": 537, "ts": 7429564856094.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856097.308, "dur": 0.139,
    "args": {
      "External id": 5966,      "cid": 370, "correlation": 2206
    }
  },
  {
    "ph": "f", "id": 2206, "pid": 537, "tid": 537, "ts": 7429564856097.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856101.902, "dur": 5.328,
    "args": {
      "External id": 5966,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2207,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2207, "pid": 537, "tid": 537, "ts": 7429564856101.902,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856107.536, "dur": 5.219,
    "args": {
      "External id": 5966,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2208,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2208, "pid": 537, "tid": 537, "ts": 7429564856107.536,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856123.600, "dur": 3.742,
    "args": {
      "External id": 5969,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2209,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2209, "pid": 537, "tid": 537, "ts": 7429564856123.600,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856157.755, "dur": 5.902,
    "args": {
      "External id": 5971,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2210,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2210, "pid": 537, "tid": 537, "ts": 7429564856157.755,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856173.483, "dur": 4.905,
    "args": {
      "External id": 5974,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2211,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2211, "pid": 537, "tid": 537, "ts": 7429564856173.483,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856186.791, "dur": 3.661,
    "args": {
      "External id": 5975,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2212,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2212, "pid": 537, "tid": 537, "ts": 7429564856186.791,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856195.717, "dur": 4.414,
    "args": {
      "External id": 5976,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2213,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2213, "pid": 537, "tid": 537, "ts": 7429564856195.717,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856207.366, "dur": 3.539,
    "args": {
      "External id": 5977,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2214,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2214, "pid": 537, "tid": 537, "ts": 7429564856207.366,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856219.029, "dur": 5.503,
    "args": {
      "External id": 5979,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2215,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2215, "pid": 537, "tid": 537, "ts": 7429564856219.029,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856263.858, "dur": 0.260,
    "args": {
      "External id": 5987,      "cid": 370, "correlation": 2216
    }
  },
  {
    "ph": "f", "id": 2216, "pid": 537, "tid": 537, "ts": 7429564856263.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856267.494, "dur": 0.156,
    "args": {
      "External id": 5987,      "cid": 370, "correlation": 2217
    }
  },
  {
    "ph": "f", "id": 2217, "pid": 537, "tid": 537, "ts": 7429564856267.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856269.822, "dur": 0.167,
    "args": {
      "External id": 5987,      "cid": 370, "correlation": 2218
    }
  },
  {
    "ph": "f", "id": 2218, "pid": 537, "tid": 537, "ts": 7429564856269.822,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856272.078, "dur": 5.991,
    "args": {
      "External id": 5987,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2219,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2219, "pid": 537, "tid": 537, "ts": 7429564856272.078,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856315.366, "dur": 0.292,
    "args": {
      "External id": 6000,      "cid": 370, "correlation": 2220
    }
  },
  {
    "ph": "f", "id": 2220, "pid": 537, "tid": 537, "ts": 7429564856315.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856318.983, "dur": 0.156,
    "args": {
      "External id": 6000,      "cid": 370, "correlation": 2221
    }
  },
  {
    "ph": "f", "id": 2221, "pid": 537, "tid": 537, "ts": 7429564856318.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856321.148, "dur": 0.145,
    "args": {
      "External id": 6000,      "cid": 370, "correlation": 2222
    }
  },
  {
    "ph": "f", "id": 2222, "pid": 537, "tid": 537, "ts": 7429564856321.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856325.594, "dur": 5.119,
    "args": {
      "External id": 6000,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2223,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2223, "pid": 537, "tid": 537, "ts": 7429564856325.594,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856331.156, "dur": 4.112,
    "args": {
      "External id": 6000,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2224,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2224, "pid": 537, "tid": 537, "ts": 7429564856331.156,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856368.165, "dur": 0.200,
    "args": {
      "External id": 6013,      "cid": 370, "correlation": 2225
    }
  },
  {
    "ph": "f", "id": 2225, "pid": 537, "tid": 537, "ts": 7429564856368.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856370.966, "dur": 0.149,
    "args": {
      "External id": 6013,      "cid": 370, "correlation": 2226
    }
  },
  {
    "ph": "f", "id": 2226, "pid": 537, "tid": 537, "ts": 7429564856370.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856372.789, "dur": 0.143,
    "args": {
      "External id": 6013,      "cid": 370, "correlation": 2227
    }
  },
  {
    "ph": "f", "id": 2227, "pid": 537, "tid": 537, "ts": 7429564856372.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856376.091, "dur": 4.610,
    "args": {
      "External id": 6013,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2228,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2228, "pid": 537, "tid": 537, "ts": 7429564856376.091,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856381.144, "dur": 4.115,
    "args": {
      "External id": 6013,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2229,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2229, "pid": 537, "tid": 537, "ts": 7429564856381.144,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856408.190, "dur": 4.909,
    "args": {
      "External id": 6023,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2230,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2230, "pid": 537, "tid": 537, "ts": 7429564856408.190,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856436.995, "dur": 5.600,
    "args": {
      "External id": 6028,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2231,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2231, "pid": 537, "tid": 537, "ts": 7429564856436.995,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856455.573, "dur": 7.300,
    "args": {
      "External id": 6029,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2232,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2232, "pid": 537, "tid": 537, "ts": 7429564856455.573,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856470.535, "dur": 4.758,
    "args": {
      "External id": 6030,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2233,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2233, "pid": 537, "tid": 537, "ts": 7429564856470.535,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856481.757, "dur": 3.571,
    "args": {
      "External id": 6031,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2234,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2234, "pid": 537, "tid": 537, "ts": 7429564856481.757,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856491.916, "dur": 4.356,
    "args": {
      "External id": 6032,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2235,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2235, "pid": 537, "tid": 537, "ts": 7429564856491.916,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856509.300, "dur": 4.052,
    "args": {
      "External id": 6037,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2236,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2236, "pid": 537, "tid": 537, "ts": 7429564856509.300,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856521.120, "dur": 4.833,
    "args": {
      "External id": 6038,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2237,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2237, "pid": 537, "tid": 537, "ts": 7429564856521.120,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856531.938, "dur": 3.183,
    "args": {
      "External id": 6039,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2238,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2238, "pid": 537, "tid": 537, "ts": 7429564856531.938,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856539.657, "dur": 4.034,
    "args": {
      "External id": 6040,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2239,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2239, "pid": 537, "tid": 537, "ts": 7429564856539.657,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856604.238, "dur": 4.303,
    "args": {
      "External id": 6057,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2240,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2240, "pid": 537, "tid": 537, "ts": 7429564856604.238,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856640.851, "dur": 3.465,
    "args": {
      "External id": 6075,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2241,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2241, "pid": 537, "tid": 537, "ts": 7429564856640.851,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856679.975, "dur": 0.329,
    "args": {
      "External id": 6088,      "cid": 370, "correlation": 2242
    }
  },
  {
    "ph": "f", "id": 2242, "pid": 537, "tid": 537, "ts": 7429564856679.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856684.278, "dur": 0.183,
    "args": {
      "External id": 6088,      "cid": 370, "correlation": 2243
    }
  },
  {
    "ph": "f", "id": 2243, "pid": 537, "tid": 537, "ts": 7429564856684.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856686.467, "dur": 0.155,
    "args": {
      "External id": 6088,      "cid": 370, "correlation": 2244
    }
  },
  {
    "ph": "f", "id": 2244, "pid": 537, "tid": 537, "ts": 7429564856686.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856689.033, "dur": 4.551,
    "args": {
      "External id": 6088,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2245,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2245, "pid": 537, "tid": 537, "ts": 7429564856689.033,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856705.855, "dur": 5.832,
    "args": {
      "External id": 6091,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 2246,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2246, "pid": 537, "tid": 537, "ts": 7429564856705.855,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856725.674, "dur": 3.761,
    "args": {
      "External id": 6098,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2247,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2247, "pid": 537, "tid": 537, "ts": 7429564856725.674,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856740.983, "dur": 5.256,
    "args": {
      "External id": 6101,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2248,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2248, "pid": 537, "tid": 537, "ts": 7429564856740.983,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856788.209, "dur": 0.220,
    "args": {
      "External id": 6113,      "cid": 370, "correlation": 2249
    }
  },
  {
    "ph": "f", "id": 2249, "pid": 537, "tid": 537, "ts": 7429564856788.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856791.641, "dur": 0.144,
    "args": {
      "External id": 6113,      "cid": 370, "correlation": 2250
    }
  },
  {
    "ph": "f", "id": 2250, "pid": 537, "tid": 537, "ts": 7429564856791.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856794.549, "dur": 0.147,
    "args": {
      "External id": 6113,      "cid": 370, "correlation": 2251
    }
  },
  {
    "ph": "f", "id": 2251, "pid": 537, "tid": 537, "ts": 7429564856794.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856796.515, "dur": 5.243,
    "args": {
      "External id": 6113,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2252,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2252, "pid": 537, "tid": 537, "ts": 7429564856796.515,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856818.555, "dur": 3.861,
    "args": {
      "External id": 6122,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2253,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2253, "pid": 537, "tid": 537, "ts": 7429564856818.555,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856861.212, "dur": 0.194,
    "args": {
      "External id": 6132,      "cid": 370, "correlation": 2254
    }
  },
  {
    "ph": "f", "id": 2254, "pid": 537, "tid": 537, "ts": 7429564856861.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856864.383, "dur": 0.152,
    "args": {
      "External id": 6132,      "cid": 370, "correlation": 2255
    }
  },
  {
    "ph": "f", "id": 2255, "pid": 537, "tid": 537, "ts": 7429564856864.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564856866.541, "dur": 0.158,
    "args": {
      "External id": 6132,      "cid": 370, "correlation": 2256
    }
  },
  {
    "ph": "f", "id": 2256, "pid": 537, "tid": 537, "ts": 7429564856866.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856868.436, "dur": 4.359,
    "args": {
      "External id": 6132,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2257,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2257, "pid": 537, "tid": 537, "ts": 7429564856868.436,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856884.879, "dur": 5.471,
    "args": {
      "External id": 6135,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2258,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2258, "pid": 537, "tid": 537, "ts": 7429564856884.879,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856907.578, "dur": 4.688,
    "args": {
      "External id": 6137,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2259,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2259, "pid": 537, "tid": 537, "ts": 7429564856907.578,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856921.606, "dur": 4.488,
    "args": {
      "External id": 6140,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2260,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2260, "pid": 537, "tid": 537, "ts": 7429564856921.606,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856934.512, "dur": 3.735,
    "args": {
      "External id": 6141,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2261,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2261, "pid": 537, "tid": 537, "ts": 7429564856934.512,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856944.980, "dur": 4.004,
    "args": {
      "External id": 6142,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2262,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2262, "pid": 537, "tid": 537, "ts": 7429564856944.980,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564856955.273, "dur": 3.348,
    "args": {
      "External id": 6143,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2263,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2263, "pid": 537, "tid": 537, "ts": 7429564856955.273,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564856999.259, "dur": 0.333,
    "args": {
      "External id": 6145,      "cid": 290, "correlation": 2264
    }
  },
  {
    "ph": "f", "id": 2264, "pid": 537, "tid": 537, "ts": 7429564856999.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564857000.037, "dur": 51.353,
    "args": {
      "External id": 6145,            "bytes": 2097152, "cid": 108, "correlation": 2265, "ptr": "0x100000002"
    }
  },
  {
    "ph": "s", "id": 2265, "pid": 537, "tid": 537, "ts": 7429564857000.037,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857054.253, "dur": 4.837,
    "args": {
      "External id": 6145,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2266,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2266, "pid": 537, "tid": 537, "ts": 7429564857054.253,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857100.494, "dur": 0.234,
    "args": {
      "External id": 6153,      "cid": 370, "correlation": 2267
    }
  },
  {
    "ph": "f", "id": 2267, "pid": 537, "tid": 537, "ts": 7429564857100.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857104.261, "dur": 0.147,
    "args": {
      "External id": 6153,      "cid": 370, "correlation": 2268
    }
  },
  {
    "ph": "f", "id": 2268, "pid": 537, "tid": 537, "ts": 7429564857104.261,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857107.056, "dur": 0.149,
    "args": {
      "External id": 6153,      "cid": 370, "correlation": 2269
    }
  },
  {
    "ph": "f", "id": 2269, "pid": 537, "tid": 537, "ts": 7429564857107.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857115.018, "dur": 6.126,
    "args": {
      "External id": 6153,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2270,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2270, "pid": 537, "tid": 537, "ts": 7429564857115.018,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857121.540, "dur": 3.802,
    "args": {
      "External id": 6153,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2271,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2271, "pid": 537, "tid": 537, "ts": 7429564857121.540,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857142.603, "dur": 4.301,
    "args": {
      "External id": 6156,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2272,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2272, "pid": 537, "tid": 537, "ts": 7429564857142.603,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564857163.738, "dur": 0.165,
    "args": {
      "External id": 6164,      "cid": 290, "correlation": 2273
    }
  },
  {
    "ph": "f", "id": 2273, "pid": 537, "tid": 537, "ts": 7429564857163.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564857164.398, "dur": 34.160,
    "args": {
      "External id": 6164,            "bytes": 2097152, "cid": 108, "correlation": 2274, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 2274, "pid": 537, "tid": 537, "ts": 7429564857164.398,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857210.273, "dur": 0.197,
    "args": {
      "External id": 6164,      "cid": 370, "correlation": 2275
    }
  },
  {
    "ph": "f", "id": 2275, "pid": 537, "tid": 537, "ts": 7429564857210.273,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857213.381, "dur": 0.181,
    "args": {
      "External id": 6164,      "cid": 370, "correlation": 2276
    }
  },
  {
    "ph": "f", "id": 2276, "pid": 537, "tid": 537, "ts": 7429564857213.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857216.141, "dur": 0.151,
    "args": {
      "External id": 6164,      "cid": 370, "correlation": 2277
    }
  },
  {
    "ph": "f", "id": 2277, "pid": 537, "tid": 537, "ts": 7429564857216.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857220.712, "dur": 5.412,
    "args": {
      "External id": 6164,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2278,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2278, "pid": 537, "tid": 537, "ts": 7429564857220.712,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857226.431, "dur": 5.142,
    "args": {
      "External id": 6164,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2279,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2279, "pid": 537, "tid": 537, "ts": 7429564857226.431,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857243.045, "dur": 4.725,
    "args": {
      "External id": 6167,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 2280,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2280, "pid": 537, "tid": 537, "ts": 7429564857243.045,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857274.260, "dur": 0.158,
    "args": {
      "External id": 6175,      "cid": 370, "correlation": 2281
    }
  },
  {
    "ph": "f", "id": 2281, "pid": 537, "tid": 537, "ts": 7429564857274.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857277.856, "dur": 0.148,
    "args": {
      "External id": 6175,      "cid": 370, "correlation": 2282
    }
  },
  {
    "ph": "f", "id": 2282, "pid": 537, "tid": 537, "ts": 7429564857277.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857279.675, "dur": 0.141,
    "args": {
      "External id": 6175,      "cid": 370, "correlation": 2283
    }
  },
  {
    "ph": "f", "id": 2283, "pid": 537, "tid": 537, "ts": 7429564857279.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857283.262, "dur": 5.082,
    "args": {
      "External id": 6175,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2284,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2284, "pid": 537, "tid": 537, "ts": 7429564857283.262,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857288.807, "dur": 2.989,
    "args": {
      "External id": 6175,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2285,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2285, "pid": 537, "tid": 537, "ts": 7429564857288.807,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857302.110, "dur": 4.100,
    "args": {
      "External id": 6178,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2286,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2286, "pid": 537, "tid": 537, "ts": 7429564857302.110,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857335.245, "dur": 4.249,
    "args": {
      "External id": 6180,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2287,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2287, "pid": 537, "tid": 537, "ts": 7429564857335.245,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857349.264, "dur": 5.434,
    "args": {
      "External id": 6183,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2288,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2288, "pid": 537, "tid": 537, "ts": 7429564857349.264,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857362.658, "dur": 5.088,
    "args": {
      "External id": 6184,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2289,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2289, "pid": 537, "tid": 537, "ts": 7429564857362.658,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857374.378, "dur": 4.363,
    "args": {
      "External id": 6185,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2290,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2290, "pid": 537, "tid": 537, "ts": 7429564857374.378,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857384.672, "dur": 3.527,
    "args": {
      "External id": 6186,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2291,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2291, "pid": 537, "tid": 537, "ts": 7429564857384.672,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857396.419, "dur": 3.455,
    "args": {
      "External id": 6188,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2292,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2292, "pid": 537, "tid": 537, "ts": 7429564857396.419,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564857428.382, "dur": 0.151,
    "args": {
      "External id": 6196,      "cid": 290, "correlation": 2293
    }
  },
  {
    "ph": "f", "id": 2293, "pid": 537, "tid": 537, "ts": 7429564857428.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564857428.823, "dur": 30.904,
    "args": {
      "External id": 6196,            "bytes": 2097152, "cid": 108, "correlation": 2294, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 2294, "pid": 537, "tid": 537, "ts": 7429564857428.823,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857472.680, "dur": 0.243,
    "args": {
      "External id": 6196,      "cid": 370, "correlation": 2295
    }
  },
  {
    "ph": "f", "id": 2295, "pid": 537, "tid": 537, "ts": 7429564857472.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857476.330, "dur": 0.163,
    "args": {
      "External id": 6196,      "cid": 370, "correlation": 2296
    }
  },
  {
    "ph": "f", "id": 2296, "pid": 537, "tid": 537, "ts": 7429564857476.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857479.026, "dur": 0.288,
    "args": {
      "External id": 6196,      "cid": 370, "correlation": 2297
    }
  },
  {
    "ph": "f", "id": 2297, "pid": 537, "tid": 537, "ts": 7429564857479.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857481.755, "dur": 4.930,
    "args": {
      "External id": 6196,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2298,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2298, "pid": 537, "tid": 537, "ts": 7429564857481.755,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857528.205, "dur": 0.236,
    "args": {
      "External id": 6209,      "cid": 370, "correlation": 2299
    }
  },
  {
    "ph": "f", "id": 2299, "pid": 537, "tid": 537, "ts": 7429564857528.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857531.524, "dur": 0.158,
    "args": {
      "External id": 6209,      "cid": 370, "correlation": 2300
    }
  },
  {
    "ph": "f", "id": 2300, "pid": 537, "tid": 537, "ts": 7429564857531.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857533.704, "dur": 0.142,
    "args": {
      "External id": 6209,      "cid": 370, "correlation": 2301
    }
  },
  {
    "ph": "f", "id": 2301, "pid": 537, "tid": 537, "ts": 7429564857533.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857538.288, "dur": 4.805,
    "args": {
      "External id": 6209,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2302,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2302, "pid": 537, "tid": 537, "ts": 7429564857538.288,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857543.554, "dur": 5.042,
    "args": {
      "External id": 6209,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2303,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2303, "pid": 537, "tid": 537, "ts": 7429564857543.554,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857583.493, "dur": 0.191,
    "args": {
      "External id": 6222,      "cid": 370, "correlation": 2304
    }
  },
  {
    "ph": "f", "id": 2304, "pid": 537, "tid": 537, "ts": 7429564857583.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857586.262, "dur": 0.144,
    "args": {
      "External id": 6222,      "cid": 370, "correlation": 2305
    }
  },
  {
    "ph": "f", "id": 2305, "pid": 537, "tid": 537, "ts": 7429564857586.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857588.509, "dur": 0.137,
    "args": {
      "External id": 6222,      "cid": 370, "correlation": 2306
    }
  },
  {
    "ph": "f", "id": 2306, "pid": 537, "tid": 537, "ts": 7429564857588.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857592.243, "dur": 4.638,
    "args": {
      "External id": 6222,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2307,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2307, "pid": 537, "tid": 537, "ts": 7429564857592.243,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857597.190, "dur": 3.126,
    "args": {
      "External id": 6222,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2308,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2308, "pid": 537, "tid": 537, "ts": 7429564857597.190,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857624.818, "dur": 4.500,
    "args": {
      "External id": 6232,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2309,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2309, "pid": 537, "tid": 537, "ts": 7429564857624.818,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857651.663, "dur": 5.891,
    "args": {
      "External id": 6237,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2310,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2310, "pid": 537, "tid": 537, "ts": 7429564857651.663,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857672.066, "dur": 7.522,
    "args": {
      "External id": 6238,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2311,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2311, "pid": 537, "tid": 537, "ts": 7429564857672.066,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857687.825, "dur": 4.841,
    "args": {
      "External id": 6239,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2312,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2312, "pid": 537, "tid": 537, "ts": 7429564857687.825,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857698.112, "dur": 4.449,
    "args": {
      "External id": 6240,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2313,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2313, "pid": 537, "tid": 537, "ts": 7429564857698.112,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857708.411, "dur": 3.885,
    "args": {
      "External id": 6241,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2314,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2314, "pid": 537, "tid": 537, "ts": 7429564857708.411,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857725.611, "dur": 3.701,
    "args": {
      "External id": 6246,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2315,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2315, "pid": 537, "tid": 537, "ts": 7429564857725.611,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857736.585, "dur": 5.647,
    "args": {
      "External id": 6247,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2316,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2316, "pid": 537, "tid": 537, "ts": 7429564857736.585,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857748.276, "dur": 3.480,
    "args": {
      "External id": 6248,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2317,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2317, "pid": 537, "tid": 537, "ts": 7429564857748.276,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857756.107, "dur": 4.245,
    "args": {
      "External id": 6249,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2318,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2318, "pid": 537, "tid": 537, "ts": 7429564857756.107,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857819.464, "dur": 6.655,
    "args": {
      "External id": 6266,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2319,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2319, "pid": 537, "tid": 537, "ts": 7429564857819.464,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857857.255, "dur": 3.405,
    "args": {
      "External id": 6284,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2320,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2320, "pid": 537, "tid": 537, "ts": 7429564857857.255,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857895.840, "dur": 0.241,
    "args": {
      "External id": 6297,      "cid": 370, "correlation": 2321
    }
  },
  {
    "ph": "f", "id": 2321, "pid": 537, "tid": 537, "ts": 7429564857895.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857900.133, "dur": 0.156,
    "args": {
      "External id": 6297,      "cid": 370, "correlation": 2322
    }
  },
  {
    "ph": "f", "id": 2322, "pid": 537, "tid": 537, "ts": 7429564857900.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564857902.694, "dur": 0.153,
    "args": {
      "External id": 6297,      "cid": 370, "correlation": 2323
    }
  },
  {
    "ph": "f", "id": 2323, "pid": 537, "tid": 537, "ts": 7429564857902.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857904.974, "dur": 4.537,
    "args": {
      "External id": 6297,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2324,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2324, "pid": 537, "tid": 537, "ts": 7429564857904.974,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857921.466, "dur": 4.859,
    "args": {
      "External id": 6300,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 2325,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2325, "pid": 537, "tid": 537, "ts": 7429564857921.466,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857940.516, "dur": 3.921,
    "args": {
      "External id": 6307,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2326,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2326, "pid": 537, "tid": 537, "ts": 7429564857940.516,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564857956.112, "dur": 6.599,
    "args": {
      "External id": 6310,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2327,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2327, "pid": 537, "tid": 537, "ts": 7429564857956.112,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858005.124, "dur": 0.255,
    "args": {
      "External id": 6322,      "cid": 370, "correlation": 2328
    }
  },
  {
    "ph": "f", "id": 2328, "pid": 537, "tid": 537, "ts": 7429564858005.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858009.326, "dur": 0.145,
    "args": {
      "External id": 6322,      "cid": 370, "correlation": 2329
    }
  },
  {
    "ph": "f", "id": 2329, "pid": 537, "tid": 537, "ts": 7429564858009.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858012.378, "dur": 0.157,
    "args": {
      "External id": 6322,      "cid": 370, "correlation": 2330
    }
  },
  {
    "ph": "f", "id": 2330, "pid": 537, "tid": 537, "ts": 7429564858012.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858014.457, "dur": 5.610,
    "args": {
      "External id": 6322,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2331,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2331, "pid": 537, "tid": 537, "ts": 7429564858014.457,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858037.664, "dur": 3.702,
    "args": {
      "External id": 6331,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2332,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2332, "pid": 537, "tid": 537, "ts": 7429564858037.664,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858112.495, "dur": 0.338,
    "args": {
      "External id": 6341,      "cid": 370, "correlation": 2333
    }
  },
  {
    "ph": "f", "id": 2333, "pid": 537, "tid": 537, "ts": 7429564858112.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858116.630, "dur": 0.160,
    "args": {
      "External id": 6341,      "cid": 370, "correlation": 2334
    }
  },
  {
    "ph": "f", "id": 2334, "pid": 537, "tid": 537, "ts": 7429564858116.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858118.840, "dur": 0.183,
    "args": {
      "External id": 6341,      "cid": 370, "correlation": 2335
    }
  },
  {
    "ph": "f", "id": 2335, "pid": 537, "tid": 537, "ts": 7429564858118.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858121.149, "dur": 5.417,
    "args": {
      "External id": 6341,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2336,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2336, "pid": 537, "tid": 537, "ts": 7429564858121.149,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858142.633, "dur": 4.547,
    "args": {
      "External id": 6344,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2337,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2337, "pid": 537, "tid": 537, "ts": 7429564858142.633,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858168.602, "dur": 5.921,
    "args": {
      "External id": 6346,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2338,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2338, "pid": 537, "tid": 537, "ts": 7429564858168.602,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858186.448, "dur": 6.088,
    "args": {
      "External id": 6349,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2339,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2339, "pid": 537, "tid": 537, "ts": 7429564858186.448,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858202.009, "dur": 4.747,
    "args": {
      "External id": 6350,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2340,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2340, "pid": 537, "tid": 537, "ts": 7429564858202.009,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858213.418, "dur": 4.250,
    "args": {
      "External id": 6351,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2341,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2341, "pid": 537, "tid": 537, "ts": 7429564858213.418,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858224.764, "dur": 5.183,
    "args": {
      "External id": 6352,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2342,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2342, "pid": 537, "tid": 537, "ts": 7429564858224.764,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858239.981, "dur": 3.917,
    "args": {
      "External id": 6354,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2343,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2343, "pid": 537, "tid": 537, "ts": 7429564858239.981,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564858270.707, "dur": 0.451,
    "args": {
      "External id": 6362,      "cid": 290, "correlation": 2344
    }
  },
  {
    "ph": "f", "id": 2344, "pid": 537, "tid": 537, "ts": 7429564858270.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564858271.632, "dur": 59.935,
    "args": {
      "External id": 6362,            "bytes": 2097152, "cid": 108, "correlation": 2345, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 2345, "pid": 537, "tid": 537, "ts": 7429564858271.632,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858344.432, "dur": 0.272,
    "args": {
      "External id": 6362,      "cid": 370, "correlation": 2346
    }
  },
  {
    "ph": "f", "id": 2346, "pid": 537, "tid": 537, "ts": 7429564858344.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858348.458, "dur": 0.148,
    "args": {
      "External id": 6362,      "cid": 370, "correlation": 2347
    }
  },
  {
    "ph": "f", "id": 2347, "pid": 537, "tid": 537, "ts": 7429564858348.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858351.217, "dur": 0.161,
    "args": {
      "External id": 6362,      "cid": 370, "correlation": 2348
    }
  },
  {
    "ph": "f", "id": 2348, "pid": 537, "tid": 537, "ts": 7429564858351.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858360.052, "dur": 5.726,
    "args": {
      "External id": 6362,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2349,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2349, "pid": 537, "tid": 537, "ts": 7429564858360.052,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858366.186, "dur": 5.195,
    "args": {
      "External id": 6362,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2350,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2350, "pid": 537, "tid": 537, "ts": 7429564858366.186,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858393.362, "dur": 5.606,
    "args": {
      "External id": 6365,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2351,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2351, "pid": 537, "tid": 537, "ts": 7429564858393.362,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564858418.780, "dur": 0.238,
    "args": {
      "External id": 6373,      "cid": 290, "correlation": 2352
    }
  },
  {
    "ph": "f", "id": 2352, "pid": 537, "tid": 537, "ts": 7429564858418.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564858419.430, "dur": 30.007,
    "args": {
      "External id": 6373,            "bytes": 2097152, "cid": 108, "correlation": 2353, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 2353, "pid": 537, "tid": 537, "ts": 7429564858419.430,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858460.121, "dur": 0.212,
    "args": {
      "External id": 6373,      "cid": 370, "correlation": 2354
    }
  },
  {
    "ph": "f", "id": 2354, "pid": 537, "tid": 537, "ts": 7429564858460.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858463.282, "dur": 0.148,
    "args": {
      "External id": 6373,      "cid": 370, "correlation": 2355
    }
  },
  {
    "ph": "f", "id": 2355, "pid": 537, "tid": 537, "ts": 7429564858463.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858465.565, "dur": 0.156,
    "args": {
      "External id": 6373,      "cid": 370, "correlation": 2356
    }
  },
  {
    "ph": "f", "id": 2356, "pid": 537, "tid": 537, "ts": 7429564858465.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858470.397, "dur": 5.643,
    "args": {
      "External id": 6373,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2357,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2357, "pid": 537, "tid": 537, "ts": 7429564858470.397,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858476.404, "dur": 4.005,
    "args": {
      "External id": 6373,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2358,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2358, "pid": 537, "tid": 537, "ts": 7429564858476.404,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858491.370, "dur": 4.639,
    "args": {
      "External id": 6376,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 2359,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2359, "pid": 537, "tid": 537, "ts": 7429564858491.370,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858522.826, "dur": 0.201,
    "args": {
      "External id": 6384,      "cid": 370, "correlation": 2360
    }
  },
  {
    "ph": "f", "id": 2360, "pid": 537, "tid": 537, "ts": 7429564858522.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858526.041, "dur": 0.149,
    "args": {
      "External id": 6384,      "cid": 370, "correlation": 2361
    }
  },
  {
    "ph": "f", "id": 2361, "pid": 537, "tid": 537, "ts": 7429564858526.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858528.281, "dur": 0.141,
    "args": {
      "External id": 6384,      "cid": 370, "correlation": 2362
    }
  },
  {
    "ph": "f", "id": 2362, "pid": 537, "tid": 537, "ts": 7429564858528.281,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858531.823, "dur": 4.700,
    "args": {
      "External id": 6384,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2363,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2363, "pid": 537, "tid": 537, "ts": 7429564858531.823,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858536.829, "dur": 3.799,
    "args": {
      "External id": 6384,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2364,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2364, "pid": 537, "tid": 537, "ts": 7429564858536.829,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858551.074, "dur": 4.994,
    "args": {
      "External id": 6387,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2365,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2365, "pid": 537, "tid": 537, "ts": 7429564858551.074,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858585.703, "dur": 4.045,
    "args": {
      "External id": 6389,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2366,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2366, "pid": 537, "tid": 537, "ts": 7429564858585.703,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858599.244, "dur": 5.481,
    "args": {
      "External id": 6392,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2367,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2367, "pid": 537, "tid": 537, "ts": 7429564858599.244,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858613.923, "dur": 3.841,
    "args": {
      "External id": 6393,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2368,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2368, "pid": 537, "tid": 537, "ts": 7429564858613.923,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858623.640, "dur": 4.196,
    "args": {
      "External id": 6394,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2369,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2369, "pid": 537, "tid": 537, "ts": 7429564858623.640,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858634.012, "dur": 4.546,
    "args": {
      "External id": 6395,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2370,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2370, "pid": 537, "tid": 537, "ts": 7429564858634.012,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858647.029, "dur": 4.597,
    "args": {
      "External id": 6397,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2371,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2371, "pid": 537, "tid": 537, "ts": 7429564858647.029,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858690.090, "dur": 0.339,
    "args": {
      "External id": 6405,      "cid": 370, "correlation": 2372
    }
  },
  {
    "ph": "f", "id": 2372, "pid": 537, "tid": 537, "ts": 7429564858690.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858693.378, "dur": 0.171,
    "args": {
      "External id": 6405,      "cid": 370, "correlation": 2373
    }
  },
  {
    "ph": "f", "id": 2373, "pid": 537, "tid": 537, "ts": 7429564858693.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858695.610, "dur": 0.176,
    "args": {
      "External id": 6405,      "cid": 370, "correlation": 2374
    }
  },
  {
    "ph": "f", "id": 2374, "pid": 537, "tid": 537, "ts": 7429564858695.610,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858697.774, "dur": 4.996,
    "args": {
      "External id": 6405,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2375,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2375, "pid": 537, "tid": 537, "ts": 7429564858697.774,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858737.719, "dur": 0.181,
    "args": {
      "External id": 6418,      "cid": 370, "correlation": 2376
    }
  },
  {
    "ph": "f", "id": 2376, "pid": 537, "tid": 537, "ts": 7429564858737.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858740.699, "dur": 0.162,
    "args": {
      "External id": 6418,      "cid": 370, "correlation": 2377
    }
  },
  {
    "ph": "f", "id": 2377, "pid": 537, "tid": 537, "ts": 7429564858740.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858742.854, "dur": 0.151,
    "args": {
      "External id": 6418,      "cid": 370, "correlation": 2378
    }
  },
  {
    "ph": "f", "id": 2378, "pid": 537, "tid": 537, "ts": 7429564858742.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858747.060, "dur": 6.803,
    "args": {
      "External id": 6418,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2379,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2379, "pid": 537, "tid": 537, "ts": 7429564858747.060,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858754.250, "dur": 3.812,
    "args": {
      "External id": 6418,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2380,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2380, "pid": 537, "tid": 537, "ts": 7429564858754.250,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858792.489, "dur": 0.151,
    "args": {
      "External id": 6431,      "cid": 370, "correlation": 2381
    }
  },
  {
    "ph": "f", "id": 2381, "pid": 537, "tid": 537, "ts": 7429564858792.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858794.847, "dur": 0.142,
    "args": {
      "External id": 6431,      "cid": 370, "correlation": 2382
    }
  },
  {
    "ph": "f", "id": 2382, "pid": 537, "tid": 537, "ts": 7429564858794.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564858796.615, "dur": 0.149,
    "args": {
      "External id": 6431,      "cid": 370, "correlation": 2383
    }
  },
  {
    "ph": "f", "id": 2383, "pid": 537, "tid": 537, "ts": 7429564858796.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858799.875, "dur": 4.764,
    "args": {
      "External id": 6431,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2384,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2384, "pid": 537, "tid": 537, "ts": 7429564858799.875,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858805.116, "dur": 3.857,
    "args": {
      "External id": 6431,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2385,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2385, "pid": 537, "tid": 537, "ts": 7429564858805.116,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858832.377, "dur": 4.347,
    "args": {
      "External id": 6441,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2386,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2386, "pid": 537, "tid": 537, "ts": 7429564858832.377,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858859.765, "dur": 4.578,
    "args": {
      "External id": 6446,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2387,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2387, "pid": 537, "tid": 537, "ts": 7429564858859.765,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858879.407, "dur": 7.843,
    "args": {
      "External id": 6447,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2388,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2388, "pid": 537, "tid": 537, "ts": 7429564858879.407,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564858895.146, "dur": 0.324,
    "args": {
      "External id": 6448,      "cid": 290, "correlation": 2389
    }
  },
  {
    "ph": "f", "id": 2389, "pid": 537, "tid": 537, "ts": 7429564858895.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564858895.810, "dur": 38.014,
    "args": {
      "External id": 6448,            "bytes": 2097152, "cid": 108, "correlation": 2390, "ptr": "0x2"
    }
  },
  {
    "ph": "s", "id": 2390, "pid": 537, "tid": 537, "ts": 7429564858895.810,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858935.953, "dur": 4.095,
    "args": {
      "External id": 6448,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2391,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2391, "pid": 537, "tid": 537, "ts": 7429564858935.953,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858947.588, "dur": 5.403,
    "args": {
      "External id": 6449,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2392,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2392, "pid": 537, "tid": 537, "ts": 7429564858947.588,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858959.485, "dur": 3.651,
    "args": {
      "External id": 6450,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2393,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2393, "pid": 537, "tid": 537, "ts": 7429564858959.485,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858978.615, "dur": 4.971,
    "args": {
      "External id": 6455,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2394,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2394, "pid": 537, "tid": 537, "ts": 7429564858978.615,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564858991.210, "dur": 4.990,
    "args": {
      "External id": 6456,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2395,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2395, "pid": 537, "tid": 537, "ts": 7429564858991.210,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859002.711, "dur": 3.893,
    "args": {
      "External id": 6457,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2396,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2396, "pid": 537, "tid": 537, "ts": 7429564859002.711,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859011.533, "dur": 3.439,
    "args": {
      "External id": 6458,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2397,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2397, "pid": 537, "tid": 537, "ts": 7429564859011.533,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859075.912, "dur": 5.133,
    "args": {
      "External id": 6475,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2398,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2398, "pid": 537, "tid": 537, "ts": 7429564859075.912,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859111.203, "dur": 4.652,
    "args": {
      "External id": 6493,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2399,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2399, "pid": 537, "tid": 537, "ts": 7429564859111.203,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859150.585, "dur": 0.337,
    "args": {
      "External id": 6506,      "cid": 370, "correlation": 2400
    }
  },
  {
    "ph": "f", "id": 2400, "pid": 537, "tid": 537, "ts": 7429564859150.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859154.806, "dur": 0.170,
    "args": {
      "External id": 6506,      "cid": 370, "correlation": 2401
    }
  },
  {
    "ph": "f", "id": 2401, "pid": 537, "tid": 537, "ts": 7429564859154.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859157.266, "dur": 0.155,
    "args": {
      "External id": 6506,      "cid": 370, "correlation": 2402
    }
  },
  {
    "ph": "f", "id": 2402, "pid": 537, "tid": 537, "ts": 7429564859157.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859160.081, "dur": 6.617,
    "args": {
      "External id": 6506,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2403,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2403, "pid": 537, "tid": 537, "ts": 7429564859160.081,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859178.428, "dur": 4.740,
    "args": {
      "External id": 6509,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 2404,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2404, "pid": 537, "tid": 537, "ts": 7429564859178.428,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859197.181, "dur": 3.869,
    "args": {
      "External id": 6516,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2405,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2405, "pid": 537, "tid": 537, "ts": 7429564859197.181,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859212.517, "dur": 6.186,
    "args": {
      "External id": 6519,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2406,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2406, "pid": 537, "tid": 537, "ts": 7429564859212.517,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859260.163, "dur": 0.228,
    "args": {
      "External id": 6531,      "cid": 370, "correlation": 2407
    }
  },
  {
    "ph": "f", "id": 2407, "pid": 537, "tid": 537, "ts": 7429564859260.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859263.361, "dur": 0.154,
    "args": {
      "External id": 6531,      "cid": 370, "correlation": 2408
    }
  },
  {
    "ph": "f", "id": 2408, "pid": 537, "tid": 537, "ts": 7429564859263.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859266.231, "dur": 0.161,
    "args": {
      "External id": 6531,      "cid": 370, "correlation": 2409
    }
  },
  {
    "ph": "f", "id": 2409, "pid": 537, "tid": 537, "ts": 7429564859266.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859268.356, "dur": 6.039,
    "args": {
      "External id": 6531,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2410,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2410, "pid": 537, "tid": 537, "ts": 7429564859268.356,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859290.058, "dur": 4.053,
    "args": {
      "External id": 6540,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2411,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2411, "pid": 537, "tid": 537, "ts": 7429564859290.058,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859336.427, "dur": 0.251,
    "args": {
      "External id": 6550,      "cid": 370, "correlation": 2412
    }
  },
  {
    "ph": "f", "id": 2412, "pid": 537, "tid": 537, "ts": 7429564859336.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859339.425, "dur": 0.155,
    "args": {
      "External id": 6550,      "cid": 370, "correlation": 2413
    }
  },
  {
    "ph": "f", "id": 2413, "pid": 537, "tid": 537, "ts": 7429564859339.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859341.497, "dur": 0.196,
    "args": {
      "External id": 6550,      "cid": 370, "correlation": 2414
    }
  },
  {
    "ph": "f", "id": 2414, "pid": 537, "tid": 537, "ts": 7429564859341.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859343.302, "dur": 5.026,
    "args": {
      "External id": 6550,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2415,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2415, "pid": 537, "tid": 537, "ts": 7429564859343.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859361.002, "dur": 4.994,
    "args": {
      "External id": 6553,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2416,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2416, "pid": 537, "tid": 537, "ts": 7429564859361.002,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859384.502, "dur": 4.616,
    "args": {
      "External id": 6555,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2417,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2417, "pid": 537, "tid": 537, "ts": 7429564859384.502,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859400.074, "dur": 6.566,
    "args": {
      "External id": 6558,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2418,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2418, "pid": 537, "tid": 537, "ts": 7429564859400.074,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859414.283, "dur": 3.865,
    "args": {
      "External id": 6559,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2419,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2419, "pid": 537, "tid": 537, "ts": 7429564859414.283,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859425.140, "dur": 5.905,
    "args": {
      "External id": 6560,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2420,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2420, "pid": 537, "tid": 537, "ts": 7429564859425.140,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859437.376, "dur": 5.079,
    "args": {
      "External id": 6561,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2421,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2421, "pid": 537, "tid": 537, "ts": 7429564859437.376,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859449.509, "dur": 3.939,
    "args": {
      "External id": 6563,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2422,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2422, "pid": 537, "tid": 537, "ts": 7429564859449.509,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859487.259, "dur": 0.190,
    "args": {
      "External id": 6571,      "cid": 370, "correlation": 2423
    }
  },
  {
    "ph": "f", "id": 2423, "pid": 537, "tid": 537, "ts": 7429564859487.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859490.264, "dur": 0.139,
    "args": {
      "External id": 6571,      "cid": 370, "correlation": 2424
    }
  },
  {
    "ph": "f", "id": 2424, "pid": 537, "tid": 537, "ts": 7429564859490.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859492.567, "dur": 0.156,
    "args": {
      "External id": 6571,      "cid": 370, "correlation": 2425
    }
  },
  {
    "ph": "f", "id": 2425, "pid": 537, "tid": 537, "ts": 7429564859492.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859499.745, "dur": 4.769,
    "args": {
      "External id": 6571,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2426,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2426, "pid": 537, "tid": 537, "ts": 7429564859499.745,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859504.918, "dur": 3.640,
    "args": {
      "External id": 6571,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2427,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2427, "pid": 537, "tid": 537, "ts": 7429564859504.918,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564859525.174, "dur": 0.339,
    "args": {
      "External id": 6574,      "cid": 290, "correlation": 2428
    }
  },
  {
    "ph": "f", "id": 2428, "pid": 537, "tid": 537, "ts": 7429564859525.174,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564859525.909, "dur": 45.646,
    "args": {
      "External id": 6574,            "bytes": 2097152, "cid": 108, "correlation": 2429, "ptr": "0x7ffcd5bf61d8"
    }
  },
  {
    "ph": "s", "id": 2429, "pid": 537, "tid": 537, "ts": 7429564859525.909,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859574.930, "dur": 6.271,
    "args": {
      "External id": 6574,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2430,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2430, "pid": 537, "tid": 537, "ts": 7429564859574.930,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859610.356, "dur": 0.210,
    "args": {
      "External id": 6582,      "cid": 370, "correlation": 2431
    }
  },
  {
    "ph": "f", "id": 2431, "pid": 537, "tid": 537, "ts": 7429564859610.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859613.528, "dur": 0.153,
    "args": {
      "External id": 6582,      "cid": 370, "correlation": 2432
    }
  },
  {
    "ph": "f", "id": 2432, "pid": 537, "tid": 537, "ts": 7429564859613.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859615.747, "dur": 0.151,
    "args": {
      "External id": 6582,      "cid": 370, "correlation": 2433
    }
  },
  {
    "ph": "f", "id": 2433, "pid": 537, "tid": 537, "ts": 7429564859615.747,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859620.850, "dur": 4.841,
    "args": {
      "External id": 6582,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2434,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2434, "pid": 537, "tid": 537, "ts": 7429564859620.850,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859626.008, "dur": 4.275,
    "args": {
      "External id": 6582,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2435,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2435, "pid": 537, "tid": 537, "ts": 7429564859626.008,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564859639.360, "dur": 0.231,
    "args": {
      "External id": 6585,      "cid": 290, "correlation": 2436
    }
  },
  {
    "ph": "f", "id": 2436, "pid": 537, "tid": 537, "ts": 7429564859639.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564859640.007, "dur": 34.423,
    "args": {
      "External id": 6585,            "bytes": 2097152, "cid": 108, "correlation": 2437, "ptr": "0x101000101000101"
    }
  },
  {
    "ph": "s", "id": 2437, "pid": 537, "tid": 537, "ts": 7429564859640.007,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859677.090, "dur": 4.340,
    "args": {
      "External id": 6585,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 2438,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2438, "pid": 537, "tid": 537, "ts": 7429564859677.090,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859710.695, "dur": 0.205,
    "args": {
      "External id": 6593,      "cid": 370, "correlation": 2439
    }
  },
  {
    "ph": "f", "id": 2439, "pid": 537, "tid": 537, "ts": 7429564859710.695,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859714.050, "dur": 0.150,
    "args": {
      "External id": 6593,      "cid": 370, "correlation": 2440
    }
  },
  {
    "ph": "f", "id": 2440, "pid": 537, "tid": 537, "ts": 7429564859714.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859716.825, "dur": 0.141,
    "args": {
      "External id": 6593,      "cid": 370, "correlation": 2441
    }
  },
  {
    "ph": "f", "id": 2441, "pid": 537, "tid": 537, "ts": 7429564859716.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859721.755, "dur": 6.132,
    "args": {
      "External id": 6593,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2442,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2442, "pid": 537, "tid": 537, "ts": 7429564859721.755,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859728.245, "dur": 3.426,
    "args": {
      "External id": 6593,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2443,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2443, "pid": 537, "tid": 537, "ts": 7429564859728.245,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859742.642, "dur": 3.811,
    "args": {
      "External id": 6596,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2444,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2444, "pid": 537, "tid": 537, "ts": 7429564859742.642,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859776.001, "dur": 4.180,
    "args": {
      "External id": 6598,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2445,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2445, "pid": 537, "tid": 537, "ts": 7429564859776.001,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859789.979, "dur": 5.670,
    "args": {
      "External id": 6601,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2446,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2446, "pid": 537, "tid": 537, "ts": 7429564859789.979,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859804.791, "dur": 3.560,
    "args": {
      "External id": 6602,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2447,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2447, "pid": 537, "tid": 537, "ts": 7429564859804.791,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859813.526, "dur": 3.827,
    "args": {
      "External id": 6603,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2448,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2448, "pid": 537, "tid": 537, "ts": 7429564859813.526,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859824.238, "dur": 4.334,
    "args": {
      "External id": 6604,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2449,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2449, "pid": 537, "tid": 537, "ts": 7429564859824.238,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859835.867, "dur": 4.451,
    "args": {
      "External id": 6606,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2450,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2450, "pid": 537, "tid": 537, "ts": 7429564859835.867,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859881.211, "dur": 0.375,
    "args": {
      "External id": 6614,      "cid": 370, "correlation": 2451
    }
  },
  {
    "ph": "f", "id": 2451, "pid": 537, "tid": 537, "ts": 7429564859881.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859884.607, "dur": 0.148,
    "args": {
      "External id": 6614,      "cid": 370, "correlation": 2452
    }
  },
  {
    "ph": "f", "id": 2452, "pid": 537, "tid": 537, "ts": 7429564859884.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859886.849, "dur": 0.155,
    "args": {
      "External id": 6614,      "cid": 370, "correlation": 2453
    }
  },
  {
    "ph": "f", "id": 2453, "pid": 537, "tid": 537, "ts": 7429564859886.849,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859889.011, "dur": 5.370,
    "args": {
      "External id": 6614,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2454,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2454, "pid": 537, "tid": 537, "ts": 7429564859889.011,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859933.115, "dur": 0.200,
    "args": {
      "External id": 6627,      "cid": 370, "correlation": 2455
    }
  },
  {
    "ph": "f", "id": 2455, "pid": 537, "tid": 537, "ts": 7429564859933.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859935.738, "dur": 0.143,
    "args": {
      "External id": 6627,      "cid": 370, "correlation": 2456
    }
  },
  {
    "ph": "f", "id": 2456, "pid": 537, "tid": 537, "ts": 7429564859935.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859937.680, "dur": 0.155,
    "args": {
      "External id": 6627,      "cid": 370, "correlation": 2457
    }
  },
  {
    "ph": "f", "id": 2457, "pid": 537, "tid": 537, "ts": 7429564859937.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859941.925, "dur": 5.673,
    "args": {
      "External id": 6627,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2458,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2458, "pid": 537, "tid": 537, "ts": 7429564859941.925,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859947.996, "dur": 3.816,
    "args": {
      "External id": 6627,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2459,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2459, "pid": 537, "tid": 537, "ts": 7429564859947.996,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859986.784, "dur": 0.216,
    "args": {
      "External id": 6640,      "cid": 370, "correlation": 2460
    }
  },
  {
    "ph": "f", "id": 2460, "pid": 537, "tid": 537, "ts": 7429564859986.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859989.297, "dur": 0.150,
    "args": {
      "External id": 6640,      "cid": 370, "correlation": 2461
    }
  },
  {
    "ph": "f", "id": 2461, "pid": 537, "tid": 537, "ts": 7429564859989.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564859991.284, "dur": 0.154,
    "args": {
      "External id": 6640,      "cid": 370, "correlation": 2462
    }
  },
  {
    "ph": "f", "id": 2462, "pid": 537, "tid": 537, "ts": 7429564859991.284,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859994.607, "dur": 4.348,
    "args": {
      "External id": 6640,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2463,      "grid": [20480, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2463, "pid": 537, "tid": 537, "ts": 7429564859994.607,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564859999.228, "dur": 3.372,
    "args": {
      "External id": 6640,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "cid": 65, "correlation": 2464,      "grid": [3328, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2464, "pid": 537, "tid": 537, "ts": 7429564859999.228,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860026.630, "dur": 5.387,
    "args": {
      "External id": 6650,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2465,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2465, "pid": 537, "tid": 537, "ts": 7429564860026.630,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860053.716, "dur": 5.831,
    "args": {
      "External id": 6655,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2466,      "grid": [104, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2466, "pid": 537, "tid": 537, "ts": 7429564860053.716,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860072.486, "dur": 7.539,
    "args": {
      "External id": 6656,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2467,      "grid": [13, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2467, "pid": 537, "tid": 537, "ts": 7429564860072.486,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860112.929, "dur": 5.719,
    "args": {
      "External id": 6657,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2468,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2468, "pid": 537, "tid": 537, "ts": 7429564860112.929,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860124.276, "dur": 5.238,
    "args": {
      "External id": 6658,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2469,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2469, "pid": 537, "tid": 537, "ts": 7429564860124.276,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860136.641, "dur": 3.501,
    "args": {
      "External id": 6659,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2470,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2470, "pid": 537, "tid": 537, "ts": 7429564860136.641,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860152.318, "dur": 3.407,
    "args": {
      "External id": 6664,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2471,      "grid": [26, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2471, "pid": 537, "tid": 537, "ts": 7429564860152.318,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860161.548, "dur": 4.604,
    "args": {
      "External id": 6665,          "kernel": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "cid": 107, "correlation": 2472,      "grid": [4, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2472, "pid": 537, "tid": 537, "ts": 7429564860161.548,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860172.035, "dur": 4.046,
    "args": {
      "External id": 6666,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2473,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2473, "pid": 537, "tid": 537, "ts": 7429564860172.035,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860181.307, "dur": 3.304,
    "args": {
      "External id": 6667,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2474,      "grid": [52, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2474, "pid": 537, "tid": 537, "ts": 7429564860181.307,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860242.253, "dur": 4.713,
    "args": {
      "External id": 6684,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2475,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2475, "pid": 537, "tid": 537, "ts": 7429564860242.253,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860275.695, "dur": 4.125,
    "args": {
      "External id": 6702,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2476,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2476, "pid": 537, "tid": 537, "ts": 7429564860275.695,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860316.783, "dur": 0.275,
    "args": {
      "External id": 6715,      "cid": 370, "correlation": 2477
    }
  },
  {
    "ph": "f", "id": 2477, "pid": 537, "tid": 537, "ts": 7429564860316.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860320.628, "dur": 0.146,
    "args": {
      "External id": 6715,      "cid": 370, "correlation": 2478
    }
  },
  {
    "ph": "f", "id": 2478, "pid": 537, "tid": 537, "ts": 7429564860320.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860322.858, "dur": 0.170,
    "args": {
      "External id": 6715,      "cid": 370, "correlation": 2479
    }
  },
  {
    "ph": "f", "id": 2479, "pid": 537, "tid": 537, "ts": 7429564860322.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860325.272, "dur": 4.450,
    "args": {
      "External id": 6715,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2480,      "grid": [8192, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2480, "pid": 537, "tid": 537, "ts": 7429564860325.272,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860341.865, "dur": 4.839,
    "args": {
      "External id": 6718,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "cid": 107, "correlation": 2481,      "grid": [6, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2481, "pid": 537, "tid": 537, "ts": 7429564860341.865,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860360.671, "dur": 4.846,
    "args": {
      "External id": 6725,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "cid": 107, "correlation": 2482,      "grid": [22, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2482, "pid": 537, "tid": 537, "ts": 7429564860360.671,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860377.216, "dur": 5.296,
    "args": {
      "External id": 6728,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2483,      "grid": [26, 1, 1],      "block": [16, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2483, "pid": 537, "tid": 537, "ts": 7429564860377.216,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860423.637, "dur": 0.215,
    "args": {
      "External id": 6740,      "cid": 370, "correlation": 2484
    }
  },
  {
    "ph": "f", "id": 2484, "pid": 537, "tid": 537, "ts": 7429564860423.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860427.089, "dur": 0.143,
    "args": {
      "External id": 6740,      "cid": 370, "correlation": 2485
    }
  },
  {
    "ph": "f", "id": 2485, "pid": 537, "tid": 537, "ts": 7429564860427.089,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860429.655, "dur": 0.148,
    "args": {
      "External id": 6740,      "cid": 370, "correlation": 2486
    }
  },
  {
    "ph": "f", "id": 2486, "pid": 537, "tid": 537, "ts": 7429564860429.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860431.726, "dur": 5.793,
    "args": {
      "External id": 6740,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2487,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2487, "pid": 537, "tid": 537, "ts": 7429564860431.726,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860453.906, "dur": 3.619,
    "args": {
      "External id": 6749,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "cid": 107, "correlation": 2488,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2488, "pid": 537, "tid": 537, "ts": 7429564860453.906,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860498.535, "dur": 0.236,
    "args": {
      "External id": 6759,      "cid": 370, "correlation": 2489
    }
  },
  {
    "ph": "f", "id": 2489, "pid": 537, "tid": 537, "ts": 7429564860498.535,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860501.233, "dur": 0.154,
    "args": {
      "External id": 6759,      "cid": 370, "correlation": 2490
    }
  },
  {
    "ph": "f", "id": 2490, "pid": 537, "tid": 537, "ts": 7429564860501.233,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860503.453, "dur": 0.163,
    "args": {
      "External id": 6759,      "cid": 370, "correlation": 2491
    }
  },
  {
    "ph": "f", "id": 2491, "pid": 537, "tid": 537, "ts": 7429564860503.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860505.217, "dur": 4.415,
    "args": {
      "External id": 6759,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2492,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2492, "pid": 537, "tid": 537, "ts": 7429564860505.217,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860521.087, "dur": 4.540,
    "args": {
      "External id": 6762,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2493,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2493, "pid": 537, "tid": 537, "ts": 7429564860521.087,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860542.270, "dur": 5.300,
    "args": {
      "External id": 6764,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2494,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2494, "pid": 537, "tid": 537, "ts": 7429564860542.270,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860556.957, "dur": 4.547,
    "args": {
      "External id": 6767,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2495,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2495, "pid": 537, "tid": 537, "ts": 7429564860556.957,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860570.215, "dur": 3.560,
    "args": {
      "External id": 6768,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2496,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2496, "pid": 537, "tid": 537, "ts": 7429564860570.215,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860579.633, "dur": 3.803,
    "args": {
      "External id": 6769,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2497,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2497, "pid": 537, "tid": 537, "ts": 7429564860579.633,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860589.609, "dur": 3.926,
    "args": {
      "External id": 6770,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2498,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2498, "pid": 537, "tid": 537, "ts": 7429564860589.609,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860601.977, "dur": 4.672,
    "args": {
      "External id": 6772,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2499,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2499, "pid": 537, "tid": 537, "ts": 7429564860601.977,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564860631.894, "dur": 0.358,
    "args": {
      "External id": 6780,      "cid": 290, "correlation": 2500
    }
  },
  {
    "ph": "f", "id": 2500, "pid": 537, "tid": 537, "ts": 7429564860631.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564860632.751, "dur": 57.391,
    "args": {
      "External id": 6780,            "bytes": 2097152, "cid": 108, "correlation": 2501, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 2501, "pid": 537, "tid": 537, "ts": 7429564860632.751,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860704.913, "dur": 0.253,
    "args": {
      "External id": 6780,      "cid": 370, "correlation": 2502
    }
  },
  {
    "ph": "f", "id": 2502, "pid": 537, "tid": 537, "ts": 7429564860704.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860708.395, "dur": 0.148,
    "args": {
      "External id": 6780,      "cid": 370, "correlation": 2503
    }
  },
  {
    "ph": "f", "id": 2503, "pid": 537, "tid": 537, "ts": 7429564860708.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860711.162, "dur": 0.144,
    "args": {
      "External id": 6780,      "cid": 370, "correlation": 2504
    }
  },
  {
    "ph": "f", "id": 2504, "pid": 537, "tid": 537, "ts": 7429564860711.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860719.227, "dur": 6.242,
    "args": {
      "External id": 6780,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2505,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2505, "pid": 537, "tid": 537, "ts": 7429564860719.227,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860726.037, "dur": 4.236,
    "args": {
      "External id": 6780,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2506,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2506, "pid": 537, "tid": 537, "ts": 7429564860726.037,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860749.134, "dur": 4.553,
    "args": {
      "External id": 6783,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2507,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2507, "pid": 537, "tid": 537, "ts": 7429564860749.134,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 537, "tid": 537,
    "ts": 7429564860773.958, "dur": 0.286,
    "args": {
      "External id": 6791,      "cid": 290, "correlation": 2508
    }
  },
  {
    "ph": "f", "id": 2508, "pid": 537, "tid": 537, "ts": 7429564860773.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 537, "tid": 537,
    "ts": 7429564860774.623, "dur": 30.574,
    "args": {
      "External id": 6791,            "bytes": 2097152, "cid": 108, "correlation": 2509, "ptr": "0x0"
    }
  },
  {
    "ph": "s", "id": 2509, "pid": 537, "tid": 537, "ts": 7429564860774.623,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860817.253, "dur": 0.237,
    "args": {
      "External id": 6791,      "cid": 370, "correlation": 2510
    }
  },
  {
    "ph": "f", "id": 2510, "pid": 537, "tid": 537, "ts": 7429564860817.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860820.408, "dur": 0.143,
    "args": {
      "External id": 6791,      "cid": 370, "correlation": 2511
    }
  },
  {
    "ph": "f", "id": 2511, "pid": 537, "tid": 537, "ts": 7429564860820.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860822.679, "dur": 0.159,
    "args": {
      "External id": 6791,      "cid": 370, "correlation": 2512
    }
  },
  {
    "ph": "f", "id": 2512, "pid": 537, "tid": 537, "ts": 7429564860822.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860827.258, "dur": 5.488,
    "args": {
      "External id": 6791,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2513,      "grid": [215040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2513, "pid": 537, "tid": 537, "ts": 7429564860827.258,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860833.244, "dur": 3.861,
    "args": {
      "External id": 6791,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2514,      "grid": [46592, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2514, "pid": 537, "tid": 537, "ts": 7429564860833.244,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860847.428, "dur": 4.546,
    "args": {
      "External id": 6794,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "cid": 107, "correlation": 2515,      "grid": [182, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2515, "pid": 537, "tid": 537, "ts": 7429564860847.428,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860881.932, "dur": 0.178,
    "args": {
      "External id": 6802,      "cid": 370, "correlation": 2516
    }
  },
  {
    "ph": "f", "id": 2516, "pid": 537, "tid": 537, "ts": 7429564860881.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860885.398, "dur": 0.141,
    "args": {
      "External id": 6802,      "cid": 370, "correlation": 2517
    }
  },
  {
    "ph": "f", "id": 2517, "pid": 537, "tid": 537, "ts": 7429564860885.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564860887.342, "dur": 0.153,
    "args": {
      "External id": 6802,      "cid": 370, "correlation": 2518
    }
  },
  {
    "ph": "f", "id": 2518, "pid": 537, "tid": 537, "ts": 7429564860887.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860890.926, "dur": 5.795,
    "args": {
      "External id": 6802,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2519,      "grid": [245760, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2519, "pid": 537, "tid": 537, "ts": 7429564860890.926,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860897.028, "dur": 3.336,
    "args": {
      "External id": 6802,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2520,      "grid": [13312, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2520, "pid": 537, "tid": 537, "ts": 7429564860897.028,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860910.556, "dur": 4.105,
    "args": {
      "External id": 6805,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "cid": 107, "correlation": 2521,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2521, "pid": 537, "tid": 537, "ts": 7429564860910.556,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860938.674, "dur": 4.439,
    "args": {
      "External id": 6807,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2522,      "grid": [52, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2522, "pid": 537, "tid": 537, "ts": 7429564860938.674,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860953.047, "dur": 5.426,
    "args": {
      "External id": 6810,          "kernel": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "cid": 107, "correlation": 2523,      "grid": [2, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2523, "pid": 537, "tid": 537, "ts": 7429564860953.047,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860965.673, "dur": 3.928,
    "args": {
      "External id": 6811,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "cid": 107, "correlation": 2524,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2524, "pid": 537, "tid": 537, "ts": 7429564860965.673,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860975.940, "dur": 5.307,
    "args": {
      "External id": 6812,          "kernel": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "cid": 107, "correlation": 2525,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2525, "pid": 537, "tid": 537, "ts": 7429564860975.940,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860987.065, "dur": 3.790,
    "args": {
      "External id": 6813,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2526,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2526, "pid": 537, "tid": 537, "ts": 7429564860987.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564860999.588, "dur": 3.522,
    "args": {
      "External id": 6815,          "kernel": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "cid": 107, "correlation": 2527,      "grid": [208, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2527, "pid": 537, "tid": 537, "ts": 7429564860999.588,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564861135.376, "dur": 0.315,
    "args": {
      "External id": 6829,      "cid": 370, "correlation": 2528
    }
  },
  {
    "ph": "f", "id": 2528, "pid": 537, "tid": 537, "ts": 7429564861135.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564861140.156, "dur": 0.176,
    "args": {
      "External id": 6829,      "cid": 370, "correlation": 2529
    }
  },
  {
    "ph": "f", "id": 2529, "pid": 537, "tid": 537, "ts": 7429564861140.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 537, "tid": 537,
    "ts": 7429564861143.200, "dur": 0.253,
    "args": {
      "External id": 6829,      "cid": 370, "correlation": 2530
    }
  },
  {
    "ph": "f", "id": 2530, "pid": 537, "tid": 537, "ts": 7429564861143.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564861149.617, "dur": 5.843,
    "args": {
      "External id": 6829,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x48x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2531,      "grid": [1539072, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2531, "pid": 537, "tid": 537, "ts": 7429564861149.617,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 537, "tid": 537,
    "ts": 7429564861155.853, "dur": 3.686,
    "args": {
      "External id": 6829,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU4_VW4", "cid": 65, "correlation": 2532,      "grid": [417024, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2532, "pid": 537, "tid": 537, "ts": 7429564861155.853,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipDeviceSynchronize", "pid": 537, "tid": 537,
    "ts": 7429564861411.478, "dur": 445.786,
    "args": {
            "cid": 48, "correlation": 2533
    }
  },
  {
    "ph": "f", "id": 2533, "pid": 537, "tid": 537, "ts": 7429564861411.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::indexSelectSmallIndex<float, long, unsigned int, 2, 2, -2>(at::cuda::detail::TensorInfo<float, unsigned int>, at::cuda::detail::TensorInfo<float const, unsigned int>, at::cuda::detail::TensorInfo<long const, unsigned int>, int, int, unsigned int, long)", "pid": 2, "tid": 0,
    "ts": 7429564823955.167, "dur": 8.819,
    "args": {
      "External id": 4,      "device": 2, "stream": 0,      "correlation": 1, "kind": "Dispatch Kernel",      "grid":     [8, 1, 1], "block":     [512, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1, "pid": 2, "tid": 0, "ts": 7429564823955.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::elementwise_kernel_with_index<int, at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}>(int, at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, function_traits<at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}>::result_type*)", "pid": 2, "tid": 0,
    "ts": 7429564823963.987, "dur": 5.372,
    "args": {
      "External id": 10,      "device": 2, "stream": 0,      "correlation": 2, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2, "pid": 2, "tid": 0, "ts": 7429564823963.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<float>, std::array<char*, 1ul> >(int, at::native::FillFunctor<float>, std::array<char*, 1ul>)", "pid": 2, "tid": 0,
    "ts": 7429564823969.360, "dur": 5.331,
    "args": {
      "External id": 16,      "device": 2, "stream": 0,      "correlation": 3, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 3, "pid": 2, "tid": 0, "ts": 7429564823969.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::triu_tril_kernel<float, int, true, 2, false>(at::cuda::detail::TensorInfo<float, int>, at::cuda::detail::TensorInfo<float const, int>, long, long, int)", "pid": 2, "tid": 0,
    "ts": 7429564823974.692, "dur": 5.210,
    "args": {
      "External id": 17,      "device": 2, "stream": 0,      "correlation": 4, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 4, "pid": 2, "tid": 0, "ts": 7429564823974.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::elementwise_kernel_with_index<int, at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}>(int, at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}, function_traits<at::native::arange_cuda_out(c10::Scalar const&, c10::Scalar const&, c10::Scalar const&, at::Tensor&)::{lambda()#1}::operator()() const::{lambda()#4}::operator()() const::{lambda(long)#1}>::result_type*)", "pid": 2, "tid": 0,
    "ts": 7429564823979.903, "dur": 5.492,
    "args": {
      "External id": 20,      "device": 2, "stream": 0,      "correlation": 5, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 5, "pid": 2, "tid": 0, "ts": 7429564823979.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast<at::native::(anonymous namespace)::CompareFunctor<long> >(at::TensorIteratorBase&, at::native::(anonymous namespace)::CompareFunctor<long> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::(anonymous namespace)::CompareFunctor<long> >(at::TensorIteratorBase&, at::native::(anonymous namespace)::CompareFunctor<long> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564823985.396, "dur": 5.451,
    "args": {
      "External id": 24,      "device": 2, "stream": 0,      "correlation": 6, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 6, "pid": 2, "tid": 0, "ts": 7429564823985.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564823990.848, "dur": 5.371,
    "args": {
      "External id": 25,      "device": 2, "stream": 0,      "correlation": 7, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [512, 1, 1]
    }
  },
  {
    "ph": "f", "id": 7, "pid": 2, "tid": 0, "ts": 7429564823990.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7429564823996.220, "dur": 5.131,
    "args": {
      "External id": 38,      "device": 2, "stream": 0,      "correlation": 8, "kind": "DtoD",      "bytes": 676, "memory bandwidth (GB/s)": 0.13172252533125486
    }
  },
  {
    "ph": "f", "id": 8, "pid": 2, "tid": 0, "ts": 7429564823996.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2}>(int, at::native::gpu_kernel_impl<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#2})", "pid": 2, "tid": 0,
    "ts": 7429564824001.352, "dur": 5.652,
    "args": {
      "External id": 54,      "device": 2, "stream": 0,      "correlation": 9, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 9, "pid": 2, "tid": 0, "ts": 7429564824001.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, bool, at::native::(anonymous namespace)::CompareEqFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, bool, at::native::(anonymous namespace)::CompareEqFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824007.005, "dur": 5.251,
    "args": {
      "External id": 55,      "device": 2, "stream": 0,      "correlation": 10, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 10, "pid": 2, "tid": 0, "ts": 7429564824007.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7429564824012.257, "dur": 5.332,
    "args": {
      "External id": 66,      "device": 2, "stream": 0,      "correlation": 11, "kind": "DtoD",      "bytes": 676, "memory bandwidth (GB/s)": 0.12675792237014813
    }
  },
  {
    "ph": "f", "id": 11, "pid": 2, "tid": 0, "ts": 7429564824012.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::masked_fill_kernel(at::TensorIterator&, c10::Scalar const&)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float, bool)#1}, std::array<char*, 3ul> >(int, at::native::(anonymous namespace)::masked_fill_kernel(at::TensorIterator&, c10::Scalar const&)::{lambda()#1}::operator()() const::{lambda()#7}::operator()() const::{lambda(float, bool)#1}, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824017.590, "dur": 5.291,
    "args": {
      "External id": 67,      "device": 2, "stream": 0,      "correlation": 12, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 12, "pid": 2, "tid": 0, "ts": 7429564824017.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7429564824022.882, "dur": 5.331,
    "args": {
      "External id": 74,      "device": 2, "stream": 0,      "correlation": 13, "kind": "DtoD",      "bytes": 676, "memory bandwidth (GB/s)": 0.12678169542385595
    }
  },
  {
    "ph": "f", "id": 13, "pid": 2, "tid": 0, "ts": 7429564824022.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<512, 1, at::native::gpu_kernel_impl<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824028.214, "dur": 5.411,
    "args": {
      "External id": 94,      "device": 2, "stream": 0,      "correlation": 14, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [512, 1, 1]
    }
  },
  {
    "ph": "f", "id": 14, "pid": 2, "tid": 0, "ts": 7429564824028.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT16x64x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG16_16_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824033.626, "dur": 6.414,
    "args": {
      "External id": 106,      "device": 2, "stream": 0,      "correlation": 18, "kind": "Dispatch Kernel",      "grid":     [256, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 18, "pid": 2, "tid": 0, "ts": 7429564824033.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 3, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564824040.041, "dur": 5.171,
    "args": {
      "External id": 111,      "device": 2, "stream": 0,      "correlation": 19, "kind": "Dispatch Kernel",      "grid":     [1, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 19, "pid": 2, "tid": 0, "ts": 7429564824040.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::cos_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::cos_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824045.213, "dur": 5.131,
    "args": {
      "External id": 112,      "device": 2, "stream": 0,      "correlation": 20, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 20, "pid": 2, "tid": 0, "ts": 7429564824045.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824050.345, "dur": 5.211,
    "args": {
      "External id": 113,      "device": 2, "stream": 0,      "correlation": 21, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 21, "pid": 2, "tid": 0, "ts": 7429564824050.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::sin_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::sin_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824055.557, "dur": 5.010,
    "args": {
      "External id": 114,      "device": 2, "stream": 0,      "correlation": 22, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 22, "pid": 2, "tid": 0, "ts": 7429564824055.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824060.568, "dur": 5.532,
    "args": {
      "External id": 115,      "device": 2, "stream": 0,      "correlation": 23, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 23, "pid": 2, "tid": 0, "ts": 7429564824060.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824066.101, "dur": 5.412,
    "args": {
      "External id": 119,      "device": 2, "stream": 0,      "correlation": 24, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 24, "pid": 2, "tid": 0, "ts": 7429564824066.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564824071.514, "dur": 8.177,
    "args": {
      "External id": 122,      "device": 2, "stream": 0,      "correlation": 25, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 25, "pid": 2, "tid": 0, "ts": 7429564824071.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824079.692, "dur": 4.971,
    "args": {
      "External id": 123,      "device": 2, "stream": 0,      "correlation": 26, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 26, "pid": 2, "tid": 0, "ts": 7429564824079.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824084.664, "dur": 5.331,
    "args": {
      "External id": 124,      "device": 2, "stream": 0,      "correlation": 27, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 27, "pid": 2, "tid": 0, "ts": 7429564824084.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824089.996, "dur": 5.251,
    "args": {
      "External id": 125,      "device": 2, "stream": 0,      "correlation": 28, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 28, "pid": 2, "tid": 0, "ts": 7429564824089.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824095.248, "dur": 5.532,
    "args": {
      "External id": 127,      "device": 2, "stream": 0,      "correlation": 29, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 29, "pid": 2, "tid": 0, "ts": 7429564824095.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824100.781, "dur": 66.752,
    "args": {
      "External id": 135,      "device": 2, "stream": 0,      "correlation": 33, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 33, "pid": 2, "tid": 0, "ts": 7429564824100.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824167.534, "dur": 15.475,
    "args": {
      "External id": 148,      "device": 2, "stream": 0,      "correlation": 37, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 37, "pid": 2, "tid": 0, "ts": 7429564824167.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564824183.010, "dur": 5.411,
    "args": {
      "External id": 148,      "device": 2, "stream": 0,      "correlation": 38, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 38, "pid": 2, "tid": 0, "ts": 7429564824183.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824188.422, "dur": 16.116,
    "args": {
      "External id": 161,      "device": 2, "stream": 0,      "correlation": 42, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 42, "pid": 2, "tid": 0, "ts": 7429564824188.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564824204.539, "dur": 5.291,
    "args": {
      "External id": 161,      "device": 2, "stream": 0,      "correlation": 43, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 43, "pid": 2, "tid": 0, "ts": 7429564824204.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824209.831, "dur": 5.612,
    "args": {
      "External id": 171,      "device": 2, "stream": 0,      "correlation": 44, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 44, "pid": 2, "tid": 0, "ts": 7429564824209.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824215.444, "dur": 5.372,
    "args": {
      "External id": 176,      "device": 2, "stream": 0,      "correlation": 45, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 45, "pid": 2, "tid": 0, "ts": 7429564824215.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564824220.817, "dur": 9.701,
    "args": {
      "External id": 177,      "device": 2, "stream": 0,      "correlation": 46, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 46, "pid": 2, "tid": 0, "ts": 7429564824220.817,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824230.519, "dur": 5.291,
    "args": {
      "External id": 178,      "device": 2, "stream": 0,      "correlation": 47, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 47, "pid": 2, "tid": 0, "ts": 7429564824230.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824235.811, "dur": 5.572,
    "args": {
      "External id": 179,      "device": 2, "stream": 0,      "correlation": 48, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 48, "pid": 2, "tid": 0, "ts": 7429564824235.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824241.384, "dur": 5.371,
    "args": {
      "External id": 180,      "device": 2, "stream": 0,      "correlation": 49, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 49, "pid": 2, "tid": 0, "ts": 7429564824241.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824246.756, "dur": 5.251,
    "args": {
      "External id": 185,      "device": 2, "stream": 0,      "correlation": 50, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 50, "pid": 2, "tid": 0, "ts": 7429564824246.756,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564824252.008, "dur": 5.211,
    "args": {
      "External id": 186,      "device": 2, "stream": 0,      "correlation": 51, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 51, "pid": 2, "tid": 0, "ts": 7429564824252.008,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824257.220, "dur": 5.532,
    "args": {
      "External id": 187,      "device": 2, "stream": 0,      "correlation": 52, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 52, "pid": 2, "tid": 0, "ts": 7429564824257.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824262.753, "dur": 5.692,
    "args": {
      "External id": 188,      "device": 2, "stream": 0,      "correlation": 53, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 53, "pid": 2, "tid": 0, "ts": 7429564824262.753,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824268.446, "dur": 5.572,
    "args": {
      "External id": 205,      "device": 2, "stream": 0,      "correlation": 54, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 54, "pid": 2, "tid": 0, "ts": 7429564824268.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824274.019, "dur": 5.893,
    "args": {
      "External id": 223,      "device": 2, "stream": 0,      "correlation": 55, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 55, "pid": 2, "tid": 0, "ts": 7429564824274.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824279.913, "dur": 5.571,
    "args": {
      "External id": 236,      "device": 2, "stream": 0,      "correlation": 59, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 59, "pid": 2, "tid": 0, "ts": 7429564824279.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824285.485, "dur": 5.412,
    "args": {
      "External id": 239,      "device": 2, "stream": 0,      "correlation": 60, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 60, "pid": 2, "tid": 0, "ts": 7429564824285.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824290.898, "dur": 5.612,
    "args": {
      "External id": 246,      "device": 2, "stream": 0,      "correlation": 61, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 61, "pid": 2, "tid": 0, "ts": 7429564824290.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564824296.511, "dur": 5.972,
    "args": {
      "External id": 249,      "device": 2, "stream": 0,      "correlation": 62, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 62, "pid": 2, "tid": 0, "ts": 7429564824296.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824302.484, "dur": 5.652,
    "args": {
      "External id": 261,      "device": 2, "stream": 0,      "correlation": 66, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 66, "pid": 2, "tid": 0, "ts": 7429564824302.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824308.137, "dur": 5.652,
    "args": {
      "External id": 270,      "device": 2, "stream": 0,      "correlation": 67, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 67, "pid": 2, "tid": 0, "ts": 7429564824308.137,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824313.790, "dur": 69.479,
    "args": {
      "External id": 280,      "device": 2, "stream": 0,      "correlation": 71, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 71, "pid": 2, "tid": 0, "ts": 7429564824313.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824383.270, "dur": 5.291,
    "args": {
      "External id": 283,      "device": 2, "stream": 0,      "correlation": 72, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 72, "pid": 2, "tid": 0, "ts": 7429564824383.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824388.562, "dur": 5.692,
    "args": {
      "External id": 285,      "device": 2, "stream": 0,      "correlation": 73, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 73, "pid": 2, "tid": 0, "ts": 7429564824388.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564824394.255, "dur": 8.018,
    "args": {
      "External id": 288,      "device": 2, "stream": 0,      "correlation": 74, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 74, "pid": 2, "tid": 0, "ts": 7429564824394.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824402.274, "dur": 5.331,
    "args": {
      "External id": 289,      "device": 2, "stream": 0,      "correlation": 75, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 75, "pid": 2, "tid": 0, "ts": 7429564824402.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824407.606, "dur": 5.331,
    "args": {
      "External id": 290,      "device": 2, "stream": 0,      "correlation": 76, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 76, "pid": 2, "tid": 0, "ts": 7429564824407.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824412.938, "dur": 5.692,
    "args": {
      "External id": 291,      "device": 2, "stream": 0,      "correlation": 77, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 77, "pid": 2, "tid": 0, "ts": 7429564824412.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824418.631, "dur": 5.813,
    "args": {
      "External id": 293,      "device": 2, "stream": 0,      "correlation": 78, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 78, "pid": 2, "tid": 0, "ts": 7429564824418.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824424.445, "dur": 75.973,
    "args": {
      "External id": 301,      "device": 2, "stream": 0,      "correlation": 82, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 82, "pid": 2, "tid": 0, "ts": 7429564824424.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564824500.419, "dur": 5.733,
    "args": {
      "External id": 301,      "device": 2, "stream": 0,      "correlation": 83, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 83, "pid": 2, "tid": 0, "ts": 7429564824500.419,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824506.153, "dur": 5.812,
    "args": {
      "External id": 304,      "device": 2, "stream": 0,      "correlation": 84, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 84, "pid": 2, "tid": 0, "ts": 7429564824506.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824511.966, "dur": 76.856,
    "args": {
      "External id": 312,      "device": 2, "stream": 0,      "correlation": 88, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 88, "pid": 2, "tid": 0, "ts": 7429564824511.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564824588.823, "dur": 5.411,
    "args": {
      "External id": 312,      "device": 2, "stream": 0,      "correlation": 89, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 89, "pid": 2, "tid": 0, "ts": 7429564824588.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824594.235, "dur": 5.893,
    "args": {
      "External id": 315,      "device": 2, "stream": 0,      "correlation": 90, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 90, "pid": 2, "tid": 0, "ts": 7429564824594.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824600.129, "dur": 88.763,
    "args": {
      "External id": 323,      "device": 2, "stream": 0,      "correlation": 94, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 94, "pid": 2, "tid": 0, "ts": 7429564824600.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564824688.893, "dur": 5.451,
    "args": {
      "External id": 323,      "device": 2, "stream": 0,      "correlation": 95, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 95, "pid": 2, "tid": 0, "ts": 7429564824688.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824694.345, "dur": 5.572,
    "args": {
      "External id": 326,      "device": 2, "stream": 0,      "correlation": 96, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 96, "pid": 2, "tid": 0, "ts": 7429564824694.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824699.918, "dur": 5.572,
    "args": {
      "External id": 328,      "device": 2, "stream": 0,      "correlation": 97, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 97, "pid": 2, "tid": 0, "ts": 7429564824699.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564824705.491, "dur": 7.576,
    "args": {
      "External id": 331,      "device": 2, "stream": 0,      "correlation": 98, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 98, "pid": 2, "tid": 0, "ts": 7429564824705.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824713.068, "dur": 5.291,
    "args": {
      "External id": 332,      "device": 2, "stream": 0,      "correlation": 99, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 99, "pid": 2, "tid": 0, "ts": 7429564824713.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824718.360, "dur": 5.412,
    "args": {
      "External id": 333,      "device": 2, "stream": 0,      "correlation": 100, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 100, "pid": 2, "tid": 0, "ts": 7429564824718.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824723.773, "dur": 5.451,
    "args": {
      "External id": 334,      "device": 2, "stream": 0,      "correlation": 101, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 101, "pid": 2, "tid": 0, "ts": 7429564824723.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824729.225, "dur": 5.933,
    "args": {
      "External id": 336,      "device": 2, "stream": 0,      "correlation": 102, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 102, "pid": 2, "tid": 0, "ts": 7429564824729.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824735.159, "dur": 69.719,
    "args": {
      "External id": 344,      "device": 2, "stream": 0,      "correlation": 106, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 106, "pid": 2, "tid": 0, "ts": 7429564824735.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824804.879, "dur": 16.116,
    "args": {
      "External id": 357,      "device": 2, "stream": 0,      "correlation": 110, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 110, "pid": 2, "tid": 0, "ts": 7429564824804.879,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564824820.996, "dur": 5.692,
    "args": {
      "External id": 357,      "device": 2, "stream": 0,      "correlation": 111, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 111, "pid": 2, "tid": 0, "ts": 7429564824820.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824826.689, "dur": 16.718,
    "args": {
      "External id": 370,      "device": 2, "stream": 0,      "correlation": 115, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 115, "pid": 2, "tid": 0, "ts": 7429564824826.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564824843.408, "dur": 5.411,
    "args": {
      "External id": 370,      "device": 2, "stream": 0,      "correlation": 116, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 116, "pid": 2, "tid": 0, "ts": 7429564824843.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824848.820, "dur": 5.652,
    "args": {
      "External id": 380,      "device": 2, "stream": 0,      "correlation": 117, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 117, "pid": 2, "tid": 0, "ts": 7429564824848.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824854.473, "dur": 5.853,
    "args": {
      "External id": 385,      "device": 2, "stream": 0,      "correlation": 118, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 118, "pid": 2, "tid": 0, "ts": 7429564824854.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564824860.327, "dur": 6.173,
    "args": {
      "External id": 386,      "device": 2, "stream": 0,      "correlation": 119, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 119, "pid": 2, "tid": 0, "ts": 7429564824860.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824866.501, "dur": 5.411,
    "args": {
      "External id": 387,      "device": 2, "stream": 0,      "correlation": 120, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 120, "pid": 2, "tid": 0, "ts": 7429564824866.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824871.913, "dur": 5.853,
    "args": {
      "External id": 388,      "device": 2, "stream": 0,      "correlation": 121, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 121, "pid": 2, "tid": 0, "ts": 7429564824871.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824877.767, "dur": 5.411,
    "args": {
      "External id": 389,      "device": 2, "stream": 0,      "correlation": 122, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 122, "pid": 2, "tid": 0, "ts": 7429564824877.767,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824883.179, "dur": 5.612,
    "args": {
      "External id": 394,      "device": 2, "stream": 0,      "correlation": 123, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 123, "pid": 2, "tid": 0, "ts": 7429564824883.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564824888.792, "dur": 5.732,
    "args": {
      "External id": 395,      "device": 2, "stream": 0,      "correlation": 124, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 124, "pid": 2, "tid": 0, "ts": 7429564824888.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824894.525, "dur": 5.532,
    "args": {
      "External id": 396,      "device": 2, "stream": 0,      "correlation": 125, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 125, "pid": 2, "tid": 0, "ts": 7429564824894.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824900.058, "dur": 5.492,
    "args": {
      "External id": 397,      "device": 2, "stream": 0,      "correlation": 126, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 126, "pid": 2, "tid": 0, "ts": 7429564824900.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824905.551, "dur": 5.652,
    "args": {
      "External id": 414,      "device": 2, "stream": 0,      "correlation": 127, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 127, "pid": 2, "tid": 0, "ts": 7429564824905.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824911.204, "dur": 5.812,
    "args": {
      "External id": 432,      "device": 2, "stream": 0,      "correlation": 128, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 128, "pid": 2, "tid": 0, "ts": 7429564824911.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824917.017, "dur": 5.492,
    "args": {
      "External id": 445,      "device": 2, "stream": 0,      "correlation": 132, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 132, "pid": 2, "tid": 0, "ts": 7429564824917.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564824922.510, "dur": 5.531,
    "args": {
      "External id": 448,      "device": 2, "stream": 0,      "correlation": 133, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 133, "pid": 2, "tid": 0, "ts": 7429564824922.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824928.042, "dur": 5.452,
    "args": {
      "External id": 455,      "device": 2, "stream": 0,      "correlation": 134, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 134, "pid": 2, "tid": 0, "ts": 7429564824928.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564824933.495, "dur": 6.093,
    "args": {
      "External id": 458,      "device": 2, "stream": 0,      "correlation": 135, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 135, "pid": 2, "tid": 0, "ts": 7429564824933.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824939.589, "dur": 5.532,
    "args": {
      "External id": 470,      "device": 2, "stream": 0,      "correlation": 139, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 139, "pid": 2, "tid": 0, "ts": 7429564824939.589,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564824945.122, "dur": 5.892,
    "args": {
      "External id": 479,      "device": 2, "stream": 0,      "correlation": 140, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 140, "pid": 2, "tid": 0, "ts": 7429564824945.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564824951.015, "dur": 69.759,
    "args": {
      "External id": 489,      "device": 2, "stream": 0,      "correlation": 144, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 144, "pid": 2, "tid": 0, "ts": 7429564824951.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825020.775, "dur": 5.492,
    "args": {
      "External id": 492,      "device": 2, "stream": 0,      "correlation": 147, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 147, "pid": 2, "tid": 0, "ts": 7429564825020.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825026.268, "dur": 5.572,
    "args": {
      "External id": 494,      "device": 2, "stream": 0,      "correlation": 148, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 148, "pid": 2, "tid": 0, "ts": 7429564825026.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564825031.841, "dur": 8.418,
    "args": {
      "External id": 497,      "device": 2, "stream": 0,      "correlation": 149, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 149, "pid": 2, "tid": 0, "ts": 7429564825031.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825040.260, "dur": 5.612,
    "args": {
      "External id": 498,      "device": 2, "stream": 0,      "correlation": 150, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 150, "pid": 2, "tid": 0, "ts": 7429564825040.260,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825045.873, "dur": 5.572,
    "args": {
      "External id": 499,      "device": 2, "stream": 0,      "correlation": 151, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 151, "pid": 2, "tid": 0, "ts": 7429564825045.873,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825051.446, "dur": 5.532,
    "args": {
      "External id": 500,      "device": 2, "stream": 0,      "correlation": 152, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 152, "pid": 2, "tid": 0, "ts": 7429564825051.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825056.979, "dur": 5.652,
    "args": {
      "External id": 502,      "device": 2, "stream": 0,      "correlation": 153, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 153, "pid": 2, "tid": 0, "ts": 7429564825056.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825062.632, "dur": 75.973,
    "args": {
      "External id": 510,      "device": 2, "stream": 0,      "correlation": 157, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 157, "pid": 2, "tid": 0, "ts": 7429564825062.632,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564825138.606, "dur": 5.532,
    "args": {
      "External id": 510,      "device": 2, "stream": 0,      "correlation": 158, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 158, "pid": 2, "tid": 0, "ts": 7429564825138.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825144.139, "dur": 5.692,
    "args": {
      "External id": 513,      "device": 2, "stream": 0,      "correlation": 159, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 159, "pid": 2, "tid": 0, "ts": 7429564825144.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825149.832, "dur": 82.830,
    "args": {
      "External id": 521,      "device": 2, "stream": 0,      "correlation": 165, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 165, "pid": 2, "tid": 0, "ts": 7429564825149.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564825232.663, "dur": 5.692,
    "args": {
      "External id": 521,      "device": 2, "stream": 0,      "correlation": 166, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 166, "pid": 2, "tid": 0, "ts": 7429564825232.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825238.356, "dur": 5.652,
    "args": {
      "External id": 524,      "device": 2, "stream": 0,      "correlation": 167, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 167, "pid": 2, "tid": 0, "ts": 7429564825238.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825244.009, "dur": 87.921,
    "args": {
      "External id": 532,      "device": 2, "stream": 0,      "correlation": 171, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 171, "pid": 2, "tid": 0, "ts": 7429564825244.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564825331.931, "dur": 5.331,
    "args": {
      "External id": 532,      "device": 2, "stream": 0,      "correlation": 172, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 172, "pid": 2, "tid": 0, "ts": 7429564825331.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825337.263, "dur": 5.411,
    "args": {
      "External id": 535,      "device": 2, "stream": 0,      "correlation": 173, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 173, "pid": 2, "tid": 0, "ts": 7429564825337.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825342.675, "dur": 6.013,
    "args": {
      "External id": 537,      "device": 2, "stream": 0,      "correlation": 174, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 174, "pid": 2, "tid": 0, "ts": 7429564825342.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564825348.689, "dur": 7.617,
    "args": {
      "External id": 540,      "device": 2, "stream": 0,      "correlation": 175, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 175, "pid": 2, "tid": 0, "ts": 7429564825348.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825356.307, "dur": 5.171,
    "args": {
      "External id": 541,      "device": 2, "stream": 0,      "correlation": 176, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 176, "pid": 2, "tid": 0, "ts": 7429564825356.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825361.479, "dur": 5.932,
    "args": {
      "External id": 542,      "device": 2, "stream": 0,      "correlation": 177, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 177, "pid": 2, "tid": 0, "ts": 7429564825361.479,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825367.412, "dur": 5.893,
    "args": {
      "External id": 543,      "device": 2, "stream": 0,      "correlation": 178, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 178, "pid": 2, "tid": 0, "ts": 7429564825367.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825373.306, "dur": 5.932,
    "args": {
      "External id": 545,      "device": 2, "stream": 0,      "correlation": 181, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 181, "pid": 2, "tid": 0, "ts": 7429564825373.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825379.239, "dur": 67.715,
    "args": {
      "External id": 553,      "device": 2, "stream": 0,      "correlation": 185, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 185, "pid": 2, "tid": 0, "ts": 7429564825379.239,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825446.955, "dur": 16.156,
    "args": {
      "External id": 566,      "device": 2, "stream": 0,      "correlation": 189, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 189, "pid": 2, "tid": 0, "ts": 7429564825446.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564825463.112, "dur": 5.732,
    "args": {
      "External id": 566,      "device": 2, "stream": 0,      "correlation": 190, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 190, "pid": 2, "tid": 0, "ts": 7429564825463.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825468.845, "dur": 16.036,
    "args": {
      "External id": 579,      "device": 2, "stream": 0,      "correlation": 194, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 194, "pid": 2, "tid": 0, "ts": 7429564825468.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564825484.882, "dur": 5.933,
    "args": {
      "External id": 579,      "device": 2, "stream": 0,      "correlation": 195, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 195, "pid": 2, "tid": 0, "ts": 7429564825484.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825490.816, "dur": 6.093,
    "args": {
      "External id": 589,      "device": 2, "stream": 0,      "correlation": 196, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 196, "pid": 2, "tid": 0, "ts": 7429564825490.816,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825496.910, "dur": 6.013,
    "args": {
      "External id": 594,      "device": 2, "stream": 0,      "correlation": 197, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 197, "pid": 2, "tid": 0, "ts": 7429564825496.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564825502.924, "dur": 5.932,
    "args": {
      "External id": 595,      "device": 2, "stream": 0,      "correlation": 198, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 198, "pid": 2, "tid": 0, "ts": 7429564825502.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825508.857, "dur": 5.773,
    "args": {
      "External id": 596,      "device": 2, "stream": 0,      "correlation": 199, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 199, "pid": 2, "tid": 0, "ts": 7429564825508.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825514.631, "dur": 5.892,
    "args": {
      "External id": 597,      "device": 2, "stream": 0,      "correlation": 200, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 200, "pid": 2, "tid": 0, "ts": 7429564825514.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825520.524, "dur": 5.452,
    "args": {
      "External id": 598,      "device": 2, "stream": 0,      "correlation": 201, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 201, "pid": 2, "tid": 0, "ts": 7429564825520.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825525.977, "dur": 5.612,
    "args": {
      "External id": 603,      "device": 2, "stream": 0,      "correlation": 202, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 202, "pid": 2, "tid": 0, "ts": 7429564825525.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564825531.590, "dur": 5.692,
    "args": {
      "External id": 604,      "device": 2, "stream": 0,      "correlation": 203, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 203, "pid": 2, "tid": 0, "ts": 7429564825531.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825537.283, "dur": 6.013,
    "args": {
      "External id": 605,      "device": 2, "stream": 0,      "correlation": 204, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 204, "pid": 2, "tid": 0, "ts": 7429564825537.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825543.297, "dur": 6.093,
    "args": {
      "External id": 606,      "device": 2, "stream": 0,      "correlation": 205, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 205, "pid": 2, "tid": 0, "ts": 7429564825543.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825549.391, "dur": 5.772,
    "args": {
      "External id": 623,      "device": 2, "stream": 0,      "correlation": 206, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 206, "pid": 2, "tid": 0, "ts": 7429564825549.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825555.164, "dur": 5.973,
    "args": {
      "External id": 641,      "device": 2, "stream": 0,      "correlation": 207, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 207, "pid": 2, "tid": 0, "ts": 7429564825555.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825561.138, "dur": 5.611,
    "args": {
      "External id": 654,      "device": 2, "stream": 0,      "correlation": 211, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 211, "pid": 2, "tid": 0, "ts": 7429564825561.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825566.750, "dur": 5.652,
    "args": {
      "External id": 657,      "device": 2, "stream": 0,      "correlation": 212, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 212, "pid": 2, "tid": 0, "ts": 7429564825566.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825572.403, "dur": 5.652,
    "args": {
      "External id": 664,      "device": 2, "stream": 0,      "correlation": 213, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 213, "pid": 2, "tid": 0, "ts": 7429564825572.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564825578.056, "dur": 6.374,
    "args": {
      "External id": 667,      "device": 2, "stream": 0,      "correlation": 214, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 214, "pid": 2, "tid": 0, "ts": 7429564825578.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825584.431, "dur": 5.853,
    "args": {
      "External id": 679,      "device": 2, "stream": 0,      "correlation": 218, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 218, "pid": 2, "tid": 0, "ts": 7429564825584.431,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825590.285, "dur": 5.852,
    "args": {
      "External id": 688,      "device": 2, "stream": 0,      "correlation": 219, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 219, "pid": 2, "tid": 0, "ts": 7429564825590.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825596.138, "dur": 67.554,
    "args": {
      "External id": 698,      "device": 2, "stream": 0,      "correlation": 223, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 223, "pid": 2, "tid": 0, "ts": 7429564825596.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825663.693, "dur": 5.933,
    "args": {
      "External id": 701,      "device": 2, "stream": 0,      "correlation": 224, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 224, "pid": 2, "tid": 0, "ts": 7429564825663.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825669.627, "dur": 5.973,
    "args": {
      "External id": 703,      "device": 2, "stream": 0,      "correlation": 225, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 225, "pid": 2, "tid": 0, "ts": 7429564825669.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564825675.601, "dur": 7.937,
    "args": {
      "External id": 706,      "device": 2, "stream": 0,      "correlation": 226, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 226, "pid": 2, "tid": 0, "ts": 7429564825675.601,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825683.539, "dur": 5.572,
    "args": {
      "External id": 707,      "device": 2, "stream": 0,      "correlation": 227, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 227, "pid": 2, "tid": 0, "ts": 7429564825683.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825689.112, "dur": 5.612,
    "args": {
      "External id": 708,      "device": 2, "stream": 0,      "correlation": 228, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 228, "pid": 2, "tid": 0, "ts": 7429564825689.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825694.725, "dur": 5.692,
    "args": {
      "External id": 709,      "device": 2, "stream": 0,      "correlation": 229, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 229, "pid": 2, "tid": 0, "ts": 7429564825694.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564825700.418, "dur": 5.932,
    "args": {
      "External id": 711,      "device": 2, "stream": 0,      "correlation": 230, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 230, "pid": 2, "tid": 0, "ts": 7429564825700.418,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825706.351, "dur": 83.070,
    "args": {
      "External id": 719,      "device": 2, "stream": 0,      "correlation": 236, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 236, "pid": 2, "tid": 0, "ts": 7429564825706.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564825789.422, "dur": 5.452,
    "args": {
      "External id": 719,      "device": 2, "stream": 0,      "correlation": 237, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 237, "pid": 2, "tid": 0, "ts": 7429564825789.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825794.875, "dur": 4.930,
    "args": {
      "External id": 722,      "device": 2, "stream": 0,      "correlation": 238, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 238, "pid": 2, "tid": 0, "ts": 7429564825794.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825799.806, "dur": 76.014,
    "args": {
      "External id": 730,      "device": 2, "stream": 0,      "correlation": 244, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 244, "pid": 2, "tid": 0, "ts": 7429564825799.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564825875.821, "dur": 5.812,
    "args": {
      "External id": 730,      "device": 2, "stream": 0,      "correlation": 245, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 245, "pid": 2, "tid": 0, "ts": 7429564825875.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825881.634, "dur": 5.973,
    "args": {
      "External id": 733,      "device": 2, "stream": 0,      "correlation": 246, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 246, "pid": 2, "tid": 0, "ts": 7429564825881.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564825887.608, "dur": 87.881,
    "args": {
      "External id": 741,      "device": 2, "stream": 0,      "correlation": 250, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 250, "pid": 2, "tid": 0, "ts": 7429564825887.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564825975.490, "dur": 5.612,
    "args": {
      "External id": 741,      "device": 2, "stream": 0,      "correlation": 251, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 251, "pid": 2, "tid": 0, "ts": 7429564825975.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825981.103, "dur": 5.612,
    "args": {
      "External id": 744,      "device": 2, "stream": 0,      "correlation": 252, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 252, "pid": 2, "tid": 0, "ts": 7429564825981.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564825986.716, "dur": 6.053,
    "args": {
      "External id": 746,      "device": 2, "stream": 0,      "correlation": 253, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 253, "pid": 2, "tid": 0, "ts": 7429564825986.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564825992.770, "dur": 8.298,
    "args": {
      "External id": 749,      "device": 2, "stream": 0,      "correlation": 254, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 254, "pid": 2, "tid": 0, "ts": 7429564825992.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826001.069, "dur": 4.489,
    "args": {
      "External id": 750,      "device": 2, "stream": 0,      "correlation": 255, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 255, "pid": 2, "tid": 0, "ts": 7429564826001.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826005.559, "dur": 4.770,
    "args": {
      "External id": 751,      "device": 2, "stream": 0,      "correlation": 256, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 256, "pid": 2, "tid": 0, "ts": 7429564826005.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826010.330, "dur": 4.770,
    "args": {
      "External id": 752,      "device": 2, "stream": 0,      "correlation": 257, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 257, "pid": 2, "tid": 0, "ts": 7429564826010.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826015.101, "dur": 5.331,
    "args": {
      "External id": 754,      "device": 2, "stream": 0,      "correlation": 258, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 258, "pid": 2, "tid": 0, "ts": 7429564826015.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564826020.433, "dur": 67.514,
    "args": {
      "External id": 762,      "device": 2, "stream": 0,      "correlation": 262, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 262, "pid": 2, "tid": 0, "ts": 7429564826020.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564826087.948, "dur": 15.876,
    "args": {
      "External id": 775,      "device": 2, "stream": 0,      "correlation": 266, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 266, "pid": 2, "tid": 0, "ts": 7429564826087.948,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564826103.825, "dur": 4.810,
    "args": {
      "External id": 775,      "device": 2, "stream": 0,      "correlation": 267, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 267, "pid": 2, "tid": 0, "ts": 7429564826103.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564826108.636, "dur": 16.397,
    "args": {
      "External id": 788,      "device": 2, "stream": 0,      "correlation": 271, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 271, "pid": 2, "tid": 0, "ts": 7429564826108.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564826125.034, "dur": 4.689,
    "args": {
      "External id": 788,      "device": 2, "stream": 0,      "correlation": 272, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 272, "pid": 2, "tid": 0, "ts": 7429564826125.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826129.724, "dur": 4.851,
    "args": {
      "External id": 798,      "device": 2, "stream": 0,      "correlation": 273, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 273, "pid": 2, "tid": 0, "ts": 7429564826129.724,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826134.576, "dur": 4.890,
    "args": {
      "External id": 803,      "device": 2, "stream": 0,      "correlation": 274, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 274, "pid": 2, "tid": 0, "ts": 7429564826134.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564826139.467, "dur": 6.013,
    "args": {
      "External id": 804,      "device": 2, "stream": 0,      "correlation": 277, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 277, "pid": 2, "tid": 0, "ts": 7429564826139.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826145.481, "dur": 4.770,
    "args": {
      "External id": 805,      "device": 2, "stream": 0,      "correlation": 278, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 278, "pid": 2, "tid": 0, "ts": 7429564826145.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826150.252, "dur": 4.730,
    "args": {
      "External id": 806,      "device": 2, "stream": 0,      "correlation": 279, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 279, "pid": 2, "tid": 0, "ts": 7429564826150.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826154.983, "dur": 4.609,
    "args": {
      "External id": 807,      "device": 2, "stream": 0,      "correlation": 280, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 280, "pid": 2, "tid": 0, "ts": 7429564826154.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826159.593, "dur": 4.610,
    "args": {
      "External id": 812,      "device": 2, "stream": 0,      "correlation": 281, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 281, "pid": 2, "tid": 0, "ts": 7429564826159.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564826164.204, "dur": 5.291,
    "args": {
      "External id": 813,      "device": 2, "stream": 0,      "correlation": 282, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 282, "pid": 2, "tid": 0, "ts": 7429564826164.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826169.496, "dur": 4.810,
    "args": {
      "External id": 814,      "device": 2, "stream": 0,      "correlation": 283, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 283, "pid": 2, "tid": 0, "ts": 7429564826169.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826174.307, "dur": 5.010,
    "args": {
      "External id": 815,      "device": 2, "stream": 0,      "correlation": 284, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 284, "pid": 2, "tid": 0, "ts": 7429564826174.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826179.318, "dur": 4.490,
    "args": {
      "External id": 832,      "device": 2, "stream": 0,      "correlation": 285, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 285, "pid": 2, "tid": 0, "ts": 7429564826179.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826192.749, "dur": 3.327,
    "args": {
      "External id": 850,      "device": 2, "stream": 0,      "correlation": 286, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 286, "pid": 2, "tid": 0, "ts": 7429564826192.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564826252.527, "dur": 5.331,
    "args": {
      "External id": 863,      "device": 2, "stream": 0,      "correlation": 290, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 290, "pid": 2, "tid": 0, "ts": 7429564826252.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826271.931, "dur": 4.249,
    "args": {
      "External id": 866,      "device": 2, "stream": 0,      "correlation": 291, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 291, "pid": 2, "tid": 0, "ts": 7429564826271.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826292.779, "dur": 2.605,
    "args": {
      "External id": 873,      "device": 2, "stream": 0,      "correlation": 292, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 292, "pid": 2, "tid": 0, "ts": 7429564826292.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564826311.502, "dur": 2.685,
    "args": {
      "External id": 876,      "device": 2, "stream": 0,      "correlation": 293, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 293, "pid": 2, "tid": 0, "ts": 7429564826311.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564826371.320, "dur": 6.013,
    "args": {
      "External id": 888,      "device": 2, "stream": 0,      "correlation": 297, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 297, "pid": 2, "tid": 0, "ts": 7429564826371.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826392.649, "dur": 2.765,
    "args": {
      "External id": 897,      "device": 2, "stream": 0,      "correlation": 298, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 298, "pid": 2, "tid": 0, "ts": 7429564826392.649,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564826450.462, "dur": 68.356,
    "args": {
      "External id": 907,      "device": 2, "stream": 0,      "correlation": 302, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 302, "pid": 2, "tid": 0, "ts": 7429564826450.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826518.819, "dur": 5.732,
    "args": {
      "External id": 910,      "device": 2, "stream": 0,      "correlation": 303, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 303, "pid": 2, "tid": 0, "ts": 7429564826518.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826524.552, "dur": 4.810,
    "args": {
      "External id": 912,      "device": 2, "stream": 0,      "correlation": 304, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 304, "pid": 2, "tid": 0, "ts": 7429564826524.552,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564826529.363, "dur": 7.697,
    "args": {
      "External id": 915,      "device": 2, "stream": 0,      "correlation": 305, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 305, "pid": 2, "tid": 0, "ts": 7429564826529.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826537.061, "dur": 4.489,
    "args": {
      "External id": 916,      "device": 2, "stream": 0,      "correlation": 306, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 306, "pid": 2, "tid": 0, "ts": 7429564826537.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826541.551, "dur": 5.933,
    "args": {
      "External id": 917,      "device": 2, "stream": 0,      "correlation": 307, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 307, "pid": 2, "tid": 0, "ts": 7429564826541.551,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826547.485, "dur": 4.208,
    "args": {
      "External id": 918,      "device": 2, "stream": 0,      "correlation": 308, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 308, "pid": 2, "tid": 0, "ts": 7429564826547.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826557.227, "dur": 4.048,
    "args": {
      "External id": 920,      "device": 2, "stream": 0,      "correlation": 309, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 309, "pid": 2, "tid": 0, "ts": 7429564826557.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564826611.352, "dur": 75.733,
    "args": {
      "External id": 928,      "device": 2, "stream": 0,      "correlation": 313, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 313, "pid": 2, "tid": 0, "ts": 7429564826611.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564826687.086, "dur": 4.329,
    "args": {
      "External id": 928,      "device": 2, "stream": 0,      "correlation": 314, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 314, "pid": 2, "tid": 0, "ts": 7429564826687.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826696.187, "dur": 1.963,
    "args": {
      "External id": 931,      "device": 2, "stream": 0,      "correlation": 317, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 317, "pid": 2, "tid": 0, "ts": 7429564826696.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564826743.014, "dur": 76.094,
    "args": {
      "External id": 939,      "device": 2, "stream": 0,      "correlation": 321, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 321, "pid": 2, "tid": 0, "ts": 7429564826743.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564826819.109, "dur": 4.850,
    "args": {
      "External id": 939,      "device": 2, "stream": 0,      "correlation": 322, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 322, "pid": 2, "tid": 0, "ts": 7429564826819.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826823.960, "dur": 4.089,
    "args": {
      "External id": 942,      "device": 2, "stream": 0,      "correlation": 325, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 325, "pid": 2, "tid": 0, "ts": 7429564826823.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564826834.665, "dur": 86.638,
    "args": {
      "External id": 950,      "device": 2, "stream": 0,      "correlation": 329, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 329, "pid": 2, "tid": 0, "ts": 7429564826834.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564826921.304, "dur": 5.532,
    "args": {
      "External id": 950,      "device": 2, "stream": 0,      "correlation": 330, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 330, "pid": 2, "tid": 0, "ts": 7429564826921.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826926.837, "dur": 4.770,
    "args": {
      "External id": 953,      "device": 2, "stream": 0,      "correlation": 331, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 331, "pid": 2, "tid": 0, "ts": 7429564826926.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826931.608, "dur": 5.171,
    "args": {
      "External id": 955,      "device": 2, "stream": 0,      "correlation": 332, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 332, "pid": 2, "tid": 0, "ts": 7429564826931.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564826936.780, "dur": 7.295,
    "args": {
      "External id": 958,      "device": 2, "stream": 0,      "correlation": 333, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 333, "pid": 2, "tid": 0, "ts": 7429564826936.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826944.076, "dur": 5.773,
    "args": {
      "External id": 959,      "device": 2, "stream": 0,      "correlation": 334, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 334, "pid": 2, "tid": 0, "ts": 7429564826944.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564826949.850, "dur": 4.930,
    "args": {
      "External id": 960,      "device": 2, "stream": 0,      "correlation": 335, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 335, "pid": 2, "tid": 0, "ts": 7429564826949.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826954.781, "dur": 4.770,
    "args": {
      "External id": 961,      "device": 2, "stream": 0,      "correlation": 336, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 336, "pid": 2, "tid": 0, "ts": 7429564826954.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564826959.552, "dur": 4.128,
    "args": {
      "External id": 963,      "device": 2, "stream": 0,      "correlation": 337, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 337, "pid": 2, "tid": 0, "ts": 7429564826959.552,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564826999.764, "dur": 68.397,
    "args": {
      "External id": 971,      "device": 2, "stream": 0,      "correlation": 341, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 341, "pid": 2, "tid": 0, "ts": 7429564826999.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564827068.162, "dur": 15.554,
    "args": {
      "External id": 984,      "device": 2, "stream": 0,      "correlation": 345, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 345, "pid": 2, "tid": 0, "ts": 7429564827068.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564827083.717, "dur": 4.049,
    "args": {
      "External id": 984,      "device": 2, "stream": 0,      "correlation": 346, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 346, "pid": 2, "tid": 0, "ts": 7429564827083.717,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564827104.004, "dur": 17.198,
    "args": {
      "External id": 997,      "device": 2, "stream": 0,      "correlation": 350, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 350, "pid": 2, "tid": 0, "ts": 7429564827104.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564827121.203, "dur": 3.929,
    "args": {
      "External id": 997,      "device": 2, "stream": 0,      "correlation": 351, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 351, "pid": 2, "tid": 0, "ts": 7429564827121.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827135.196, "dur": 3.687,
    "args": {
      "External id": 1007,      "device": 2, "stream": 0,      "correlation": 352, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 352, "pid": 2, "tid": 0, "ts": 7429564827135.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827162.218, "dur": 3.407,
    "args": {
      "External id": 1012,      "device": 2, "stream": 0,      "correlation": 353, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 353, "pid": 2, "tid": 0, "ts": 7429564827162.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564827185.150, "dur": 7.056,
    "args": {
      "External id": 1013,      "device": 2, "stream": 0,      "correlation": 354, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 354, "pid": 2, "tid": 0, "ts": 7429564827185.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827196.296, "dur": 3.487,
    "args": {
      "External id": 1014,      "device": 2, "stream": 0,      "correlation": 355, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 355, "pid": 2, "tid": 0, "ts": 7429564827196.296,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827208.043, "dur": 3.567,
    "args": {
      "External id": 1015,      "device": 2, "stream": 0,      "correlation": 356, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 356, "pid": 2, "tid": 0, "ts": 7429564827208.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827216.743, "dur": 3.688,
    "args": {
      "External id": 1016,      "device": 2, "stream": 0,      "correlation": 357, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 357, "pid": 2, "tid": 0, "ts": 7429564827216.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827234.063, "dur": 3.687,
    "args": {
      "External id": 1021,      "device": 2, "stream": 0,      "correlation": 358, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 358, "pid": 2, "tid": 0, "ts": 7429564827234.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564827245.209, "dur": 5.732,
    "args": {
      "External id": 1022,      "device": 2, "stream": 0,      "correlation": 359, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 359, "pid": 2, "tid": 0, "ts": 7429564827245.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827256.434, "dur": 2.926,
    "args": {
      "External id": 1023,      "device": 2, "stream": 0,      "correlation": 360, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 360, "pid": 2, "tid": 0, "ts": 7429564827256.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827263.891, "dur": 3.808,
    "args": {
      "External id": 1024,      "device": 2, "stream": 0,      "correlation": 361, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 361, "pid": 2, "tid": 0, "ts": 7429564827263.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827360.714, "dur": 2.966,
    "args": {
      "External id": 1041,      "device": 2, "stream": 0,      "correlation": 362, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 362, "pid": 2, "tid": 0, "ts": 7429564827360.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827393.389, "dur": 2.765,
    "args": {
      "External id": 1059,      "device": 2, "stream": 0,      "correlation": 363, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 363, "pid": 2, "tid": 0, "ts": 7429564827393.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564827444.266, "dur": 5.251,
    "args": {
      "External id": 1072,      "device": 2, "stream": 0,      "correlation": 367, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 367, "pid": 2, "tid": 0, "ts": 7429564827444.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564827460.824, "dur": 3.287,
    "args": {
      "External id": 1075,      "device": 2, "stream": 0,      "correlation": 368, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 368, "pid": 2, "tid": 0, "ts": 7429564827460.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827479.026, "dur": 3.126,
    "args": {
      "External id": 1082,      "device": 2, "stream": 0,      "correlation": 369, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 369, "pid": 2, "tid": 0, "ts": 7429564827479.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564827496.586, "dur": 2.966,
    "args": {
      "External id": 1085,      "device": 2, "stream": 0,      "correlation": 370, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 370, "pid": 2, "tid": 0, "ts": 7429564827496.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564827554.199, "dur": 4.850,
    "args": {
      "External id": 1097,      "device": 2, "stream": 0,      "correlation": 374, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 374, "pid": 2, "tid": 0, "ts": 7429564827554.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827573.643, "dur": 3.207,
    "args": {
      "External id": 1106,      "device": 2, "stream": 0,      "correlation": 375, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 375, "pid": 2, "tid": 0, "ts": 7429564827573.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564827629.572, "dur": 67.715,
    "args": {
      "External id": 1116,      "device": 2, "stream": 0,      "correlation": 379, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 379, "pid": 2, "tid": 0, "ts": 7429564827629.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564827697.288, "dur": 4.569,
    "args": {
      "External id": 1119,      "device": 2, "stream": 0,      "correlation": 380, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 380, "pid": 2, "tid": 0, "ts": 7429564827697.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564827701.858, "dur": 4.570,
    "args": {
      "External id": 1121,      "device": 2, "stream": 0,      "correlation": 381, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 381, "pid": 2, "tid": 0, "ts": 7429564827701.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564827706.429, "dur": 7.015,
    "args": {
      "External id": 1124,      "device": 2, "stream": 0,      "correlation": 382, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 382, "pid": 2, "tid": 0, "ts": 7429564827706.429,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564827713.445, "dur": 4.529,
    "args": {
      "External id": 1125,      "device": 2, "stream": 0,      "correlation": 383, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 383, "pid": 2, "tid": 0, "ts": 7429564827713.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564827717.975, "dur": 4.931,
    "args": {
      "External id": 1126,      "device": 2, "stream": 0,      "correlation": 384, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 384, "pid": 2, "tid": 0, "ts": 7429564827717.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827722.907, "dur": 4.208,
    "args": {
      "External id": 1127,      "device": 2, "stream": 0,      "correlation": 385, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 385, "pid": 2, "tid": 0, "ts": 7429564827722.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564827798.921, "dur": 3.447,
    "args": {
      "External id": 1129,      "device": 2, "stream": 0,      "correlation": 388, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 388, "pid": 2, "tid": 0, "ts": 7429564827798.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564827867.118, "dur": 77.297,
    "args": {
      "External id": 1137,      "device": 2, "stream": 0,      "correlation": 392, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 392, "pid": 2, "tid": 0, "ts": 7429564827867.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564827944.416, "dur": 4.810,
    "args": {
      "External id": 1137,      "device": 2, "stream": 0,      "correlation": 393, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 393, "pid": 2, "tid": 0, "ts": 7429564827944.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564827949.227, "dur": 4.168,
    "args": {
      "External id": 1140,      "device": 2, "stream": 0,      "correlation": 394, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 394, "pid": 2, "tid": 0, "ts": 7429564827949.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564827977.131, "dur": 77.176,
    "args": {
      "External id": 1148,      "device": 2, "stream": 0,      "correlation": 400, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 400, "pid": 2, "tid": 0, "ts": 7429564827977.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564828054.308, "dur": 4.810,
    "args": {
      "External id": 1148,      "device": 2, "stream": 0,      "correlation": 401, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 401, "pid": 2, "tid": 0, "ts": 7429564828054.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564828059.119, "dur": 4.931,
    "args": {
      "External id": 1151,      "device": 2, "stream": 0,      "correlation": 402, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 402, "pid": 2, "tid": 0, "ts": 7429564828059.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564828064.051, "dur": 88.121,
    "args": {
      "External id": 1159,      "device": 2, "stream": 0,      "correlation": 406, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 406, "pid": 2, "tid": 0, "ts": 7429564828064.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564828152.173, "dur": 4.690,
    "args": {
      "External id": 1159,      "device": 2, "stream": 0,      "correlation": 407, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 407, "pid": 2, "tid": 0, "ts": 7429564828152.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564828156.864, "dur": 4.610,
    "args": {
      "External id": 1162,      "device": 2, "stream": 0,      "correlation": 408, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 408, "pid": 2, "tid": 0, "ts": 7429564828156.864,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564828161.475, "dur": 4.689,
    "args": {
      "External id": 1164,      "device": 2, "stream": 0,      "correlation": 409, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 409, "pid": 2, "tid": 0, "ts": 7429564828161.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564828166.165, "dur": 7.457,
    "args": {
      "External id": 1167,      "device": 2, "stream": 0,      "correlation": 410, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 410, "pid": 2, "tid": 0, "ts": 7429564828166.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564828173.623, "dur": 4.529,
    "args": {
      "External id": 1168,      "device": 2, "stream": 0,      "correlation": 411, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 411, "pid": 2, "tid": 0, "ts": 7429564828173.623,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564828178.153, "dur": 4.650,
    "args": {
      "External id": 1169,      "device": 2, "stream": 0,      "correlation": 412, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 412, "pid": 2, "tid": 0, "ts": 7429564828178.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828182.804, "dur": 4.689,
    "args": {
      "External id": 1170,      "device": 2, "stream": 0,      "correlation": 413, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 413, "pid": 2, "tid": 0, "ts": 7429564828182.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828187.494, "dur": 4.570,
    "args": {
      "External id": 1172,      "device": 2, "stream": 0,      "correlation": 414, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 414, "pid": 2, "tid": 0, "ts": 7429564828187.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564828246.630, "dur": 67.635,
    "args": {
      "External id": 1180,      "device": 2, "stream": 0,      "correlation": 420, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 420, "pid": 2, "tid": 0, "ts": 7429564828246.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564828314.266, "dur": 15.996,
    "args": {
      "External id": 1193,      "device": 2, "stream": 0,      "correlation": 424, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 424, "pid": 2, "tid": 0, "ts": 7429564828314.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564828330.263, "dur": 4.008,
    "args": {
      "External id": 1193,      "device": 2, "stream": 0,      "correlation": 425, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 425, "pid": 2, "tid": 0, "ts": 7429564828330.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564828362.617, "dur": 15.795,
    "args": {
      "External id": 1206,      "device": 2, "stream": 0,      "correlation": 429, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 429, "pid": 2, "tid": 0, "ts": 7429564828362.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564828378.413, "dur": 3.968,
    "args": {
      "External id": 1206,      "device": 2, "stream": 0,      "correlation": 430, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 430, "pid": 2, "tid": 0, "ts": 7429564828378.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828395.452, "dur": 2.605,
    "args": {
      "External id": 1216,      "device": 2, "stream": 0,      "correlation": 431, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 431, "pid": 2, "tid": 0, "ts": 7429564828395.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828422.475, "dur": 2.925,
    "args": {
      "External id": 1221,      "device": 2, "stream": 0,      "correlation": 432, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 432, "pid": 2, "tid": 0, "ts": 7429564828422.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564828443.844, "dur": 5.692,
    "args": {
      "External id": 1222,      "device": 2, "stream": 0,      "correlation": 433, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 433, "pid": 2, "tid": 0, "ts": 7429564828443.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828453.626, "dur": 3.888,
    "args": {
      "External id": 1223,      "device": 2, "stream": 0,      "correlation": 434, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 434, "pid": 2, "tid": 0, "ts": 7429564828453.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828465.734, "dur": 2.765,
    "args": {
      "External id": 1224,      "device": 2, "stream": 0,      "correlation": 435, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 435, "pid": 2, "tid": 0, "ts": 7429564828465.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828475.236, "dur": 2.444,
    "args": {
      "External id": 1225,      "device": 2, "stream": 0,      "correlation": 436, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 436, "pid": 2, "tid": 0, "ts": 7429564828475.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828492.997, "dur": 2.725,
    "args": {
      "External id": 1230,      "device": 2, "stream": 0,      "correlation": 437, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 437, "pid": 2, "tid": 0, "ts": 7429564828492.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564828503.781, "dur": 5.893,
    "args": {
      "External id": 1231,      "device": 2, "stream": 0,      "correlation": 438, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 438, "pid": 2, "tid": 0, "ts": 7429564828503.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828513.484, "dur": 3.767,
    "args": {
      "External id": 1232,      "device": 2, "stream": 0,      "correlation": 439, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 439, "pid": 2, "tid": 0, "ts": 7429564828513.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828521.783, "dur": 3.808,
    "args": {
      "External id": 1233,      "device": 2, "stream": 0,      "correlation": 440, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 440, "pid": 2, "tid": 0, "ts": 7429564828521.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828586.291, "dur": 2.765,
    "args": {
      "External id": 1250,      "device": 2, "stream": 0,      "correlation": 441, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 441, "pid": 2, "tid": 0, "ts": 7429564828586.291,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828617.683, "dur": 3.327,
    "args": {
      "External id": 1268,      "device": 2, "stream": 0,      "correlation": 442, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 442, "pid": 2, "tid": 0, "ts": 7429564828617.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564828666.475, "dur": 5.733,
    "args": {
      "External id": 1281,      "device": 2, "stream": 0,      "correlation": 446, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 446, "pid": 2, "tid": 0, "ts": 7429564828666.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564828682.232, "dur": 3.366,
    "args": {
      "External id": 1284,      "device": 2, "stream": 0,      "correlation": 447, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 447, "pid": 2, "tid": 0, "ts": 7429564828682.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828699.672, "dur": 2.885,
    "args": {
      "External id": 1291,      "device": 2, "stream": 0,      "correlation": 448, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 448, "pid": 2, "tid": 0, "ts": 7429564828699.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564828718.114, "dur": 2.485,
    "args": {
      "External id": 1294,      "device": 2, "stream": 0,      "correlation": 449, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 449, "pid": 2, "tid": 0, "ts": 7429564828718.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564828771.156, "dur": 5.532,
    "args": {
      "External id": 1306,      "device": 2, "stream": 0,      "correlation": 453, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 453, "pid": 2, "tid": 0, "ts": 7429564828771.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828792.244, "dur": 2.726,
    "args": {
      "External id": 1315,      "device": 2, "stream": 0,      "correlation": 454, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 454, "pid": 2, "tid": 0, "ts": 7429564828792.244,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564828844.966, "dur": 68.155,
    "args": {
      "External id": 1325,      "device": 2, "stream": 0,      "correlation": 458, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 458, "pid": 2, "tid": 0, "ts": 7429564828844.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564828913.122, "dur": 4.730,
    "args": {
      "External id": 1328,      "device": 2, "stream": 0,      "correlation": 459, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 459, "pid": 2, "tid": 0, "ts": 7429564828913.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564828917.853, "dur": 5.412,
    "args": {
      "External id": 1330,      "device": 2, "stream": 0,      "correlation": 460, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 460, "pid": 2, "tid": 0, "ts": 7429564828917.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564828923.266, "dur": 6.854,
    "args": {
      "External id": 1333,      "device": 2, "stream": 0,      "correlation": 461, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 461, "pid": 2, "tid": 0, "ts": 7429564828923.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564828930.121, "dur": 4.450,
    "args": {
      "External id": 1334,      "device": 2, "stream": 0,      "correlation": 462, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 462, "pid": 2, "tid": 0, "ts": 7429564828930.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564828934.572, "dur": 4.569,
    "args": {
      "External id": 1335,      "device": 2, "stream": 0,      "correlation": 463, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 463, "pid": 2, "tid": 0, "ts": 7429564828934.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828939.142, "dur": 4.209,
    "args": {
      "External id": 1336,      "device": 2, "stream": 0,      "correlation": 464, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 464, "pid": 2, "tid": 0, "ts": 7429564828939.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564828948.083, "dur": 2.444,
    "args": {
      "External id": 1338,      "device": 2, "stream": 0,      "correlation": 465, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 465, "pid": 2, "tid": 0, "ts": 7429564828948.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564829061.624, "dur": 78.339,
    "args": {
      "External id": 1346,      "device": 2, "stream": 0,      "correlation": 471, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 471, "pid": 2, "tid": 0, "ts": 7429564829061.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564829139.964, "dur": 4.609,
    "args": {
      "External id": 1346,      "device": 2, "stream": 0,      "correlation": 472, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 472, "pid": 2, "tid": 0, "ts": 7429564829139.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564829144.574, "dur": 4.089,
    "args": {
      "External id": 1349,      "device": 2, "stream": 0,      "correlation": 473, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 473, "pid": 2, "tid": 0, "ts": 7429564829144.574,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564829173.280, "dur": 75.774,
    "args": {
      "External id": 1357,      "device": 2, "stream": 0,      "correlation": 479, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 479, "pid": 2, "tid": 0, "ts": 7429564829173.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564829249.055, "dur": 4.930,
    "args": {
      "External id": 1357,      "device": 2, "stream": 0,      "correlation": 480, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 480, "pid": 2, "tid": 0, "ts": 7429564829249.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564829253.986, "dur": 5.411,
    "args": {
      "External id": 1360,      "device": 2, "stream": 0,      "correlation": 481, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 481, "pid": 2, "tid": 0, "ts": 7429564829253.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564829259.398, "dur": 89.204,
    "args": {
      "External id": 1368,      "device": 2, "stream": 0,      "correlation": 485, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 485, "pid": 2, "tid": 0, "ts": 7429564829259.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564829348.603, "dur": 5.813,
    "args": {
      "External id": 1368,      "device": 2, "stream": 0,      "correlation": 486, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 486, "pid": 2, "tid": 0, "ts": 7429564829348.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564829354.417, "dur": 4.649,
    "args": {
      "External id": 1371,      "device": 2, "stream": 0,      "correlation": 487, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 487, "pid": 2, "tid": 0, "ts": 7429564829354.417,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564829359.067, "dur": 4.770,
    "args": {
      "External id": 1373,      "device": 2, "stream": 0,      "correlation": 488, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 488, "pid": 2, "tid": 0, "ts": 7429564829359.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564829363.838, "dur": 6.855,
    "args": {
      "External id": 1376,      "device": 2, "stream": 0,      "correlation": 489, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 489, "pid": 2, "tid": 0, "ts": 7429564829363.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564829370.694, "dur": 4.449,
    "args": {
      "External id": 1377,      "device": 2, "stream": 0,      "correlation": 490, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 490, "pid": 2, "tid": 0, "ts": 7429564829370.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564829375.144, "dur": 4.409,
    "args": {
      "External id": 1378,      "device": 2, "stream": 0,      "correlation": 491, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 491, "pid": 2, "tid": 0, "ts": 7429564829375.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829379.554, "dur": 4.730,
    "args": {
      "External id": 1379,      "device": 2, "stream": 0,      "correlation": 492, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 492, "pid": 2, "tid": 0, "ts": 7429564829379.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829384.285, "dur": 4.009,
    "args": {
      "External id": 1381,      "device": 2, "stream": 0,      "correlation": 493, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 493, "pid": 2, "tid": 0, "ts": 7429564829384.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564829407.258, "dur": 67.795,
    "args": {
      "External id": 1389,      "device": 2, "stream": 0,      "correlation": 497, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 497, "pid": 2, "tid": 0, "ts": 7429564829407.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564829475.054, "dur": 15.434,
    "args": {
      "External id": 1402,      "device": 2, "stream": 0,      "correlation": 501, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 501, "pid": 2, "tid": 0, "ts": 7429564829475.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564829490.489, "dur": 3.888,
    "args": {
      "External id": 1402,      "device": 2, "stream": 0,      "correlation": 502, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 502, "pid": 2, "tid": 0, "ts": 7429564829490.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564829513.061, "dur": 17.480,
    "args": {
      "External id": 1415,      "device": 2, "stream": 0,      "correlation": 506, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 506, "pid": 2, "tid": 0, "ts": 7429564829513.061,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564829530.542, "dur": 4.008,
    "args": {
      "External id": 1415,      "device": 2, "stream": 0,      "correlation": 507, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 507, "pid": 2, "tid": 0, "ts": 7429564829530.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829545.937, "dur": 4.088,
    "args": {
      "External id": 1425,      "device": 2, "stream": 0,      "correlation": 508, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 508, "pid": 2, "tid": 0, "ts": 7429564829545.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829573.400, "dur": 3.527,
    "args": {
      "External id": 1430,      "device": 2, "stream": 0,      "correlation": 509, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 509, "pid": 2, "tid": 0, "ts": 7429564829573.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564829595.451, "dur": 6.694,
    "args": {
      "External id": 1431,      "device": 2, "stream": 0,      "correlation": 510, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 510, "pid": 2, "tid": 0, "ts": 7429564829595.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829646.247, "dur": 3.808,
    "args": {
      "External id": 1432,      "device": 2, "stream": 0,      "correlation": 513, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 513, "pid": 2, "tid": 0, "ts": 7429564829646.247,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829661.162, "dur": 3.527,
    "args": {
      "External id": 1433,      "device": 2, "stream": 0,      "correlation": 514, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 514, "pid": 2, "tid": 0, "ts": 7429564829661.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829670.102, "dur": 3.808,
    "args": {
      "External id": 1434,      "device": 2, "stream": 0,      "correlation": 515, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 515, "pid": 2, "tid": 0, "ts": 7429564829670.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829688.745, "dur": 3.567,
    "args": {
      "External id": 1439,      "device": 2, "stream": 0,      "correlation": 516, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 516, "pid": 2, "tid": 0, "ts": 7429564829688.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564829701.254, "dur": 6.333,
    "args": {
      "External id": 1440,      "device": 2, "stream": 0,      "correlation": 517, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 517, "pid": 2, "tid": 0, "ts": 7429564829701.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829711.638, "dur": 3.567,
    "args": {
      "External id": 1441,      "device": 2, "stream": 0,      "correlation": 518, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 518, "pid": 2, "tid": 0, "ts": 7429564829711.638,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829719.817, "dur": 3.847,
    "args": {
      "External id": 1442,      "device": 2, "stream": 0,      "correlation": 519, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 519, "pid": 2, "tid": 0, "ts": 7429564829719.817,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829786.129, "dur": 3.768,
    "args": {
      "External id": 1459,      "device": 2, "stream": 0,      "correlation": 520, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 520, "pid": 2, "tid": 0, "ts": 7429564829786.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829819.726, "dur": 3.808,
    "args": {
      "External id": 1477,      "device": 2, "stream": 0,      "correlation": 521, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 521, "pid": 2, "tid": 0, "ts": 7429564829819.726,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564829874.853, "dur": 6.013,
    "args": {
      "External id": 1490,      "device": 2, "stream": 0,      "correlation": 525, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 525, "pid": 2, "tid": 0, "ts": 7429564829874.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564829891.331, "dur": 3.848,
    "args": {
      "External id": 1493,      "device": 2, "stream": 0,      "correlation": 526, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 526, "pid": 2, "tid": 0, "ts": 7429564829891.331,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564829912.018, "dur": 2.886,
    "args": {
      "External id": 1500,      "device": 2, "stream": 0,      "correlation": 527, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 527, "pid": 2, "tid": 0, "ts": 7429564829912.018,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564829931.704, "dur": 2.685,
    "args": {
      "External id": 1503,      "device": 2, "stream": 0,      "correlation": 528, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 528, "pid": 2, "tid": 0, "ts": 7429564829931.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564829989.035, "dur": 5.091,
    "args": {
      "External id": 1515,      "device": 2, "stream": 0,      "correlation": 532, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 532, "pid": 2, "tid": 0, "ts": 7429564829989.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830009.563, "dur": 3.046,
    "args": {
      "External id": 1524,      "device": 2, "stream": 0,      "correlation": 533, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 533, "pid": 2, "tid": 0, "ts": 7429564830009.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564830063.005, "dur": 67.354,
    "args": {
      "External id": 1534,      "device": 2, "stream": 0,      "correlation": 537, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 537, "pid": 2, "tid": 0, "ts": 7429564830063.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564830135.452, "dur": 2.445,
    "args": {
      "External id": 1537,      "device": 2, "stream": 0,      "correlation": 538, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 538, "pid": 2, "tid": 0, "ts": 7429564830135.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564830163.677, "dur": 2.966,
    "args": {
      "External id": 1539,      "device": 2, "stream": 0,      "correlation": 539, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 539, "pid": 2, "tid": 0, "ts": 7429564830163.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564830183.322, "dur": 6.895,
    "args": {
      "External id": 1542,      "device": 2, "stream": 0,      "correlation": 540, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 540, "pid": 2, "tid": 0, "ts": 7429564830183.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564830194.949, "dur": 3.567,
    "args": {
      "External id": 1543,      "device": 2, "stream": 0,      "correlation": 541, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 541, "pid": 2, "tid": 0, "ts": 7429564830194.949,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564830207.578, "dur": 2.765,
    "args": {
      "External id": 1544,      "device": 2, "stream": 0,      "correlation": 542, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 542, "pid": 2, "tid": 0, "ts": 7429564830207.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830218.082, "dur": 2.886,
    "args": {
      "External id": 1545,      "device": 2, "stream": 0,      "correlation": 543, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 543, "pid": 2, "tid": 0, "ts": 7429564830218.082,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830231.593, "dur": 3.086,
    "args": {
      "External id": 1547,      "device": 2, "stream": 0,      "correlation": 544, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 544, "pid": 2, "tid": 0, "ts": 7429564830231.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564830291.972, "dur": 80.905,
    "args": {
      "External id": 1555,      "device": 2, "stream": 0,      "correlation": 548, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 548, "pid": 2, "tid": 0, "ts": 7429564830291.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564830372.878, "dur": 6.373,
    "args": {
      "External id": 1555,      "device": 2, "stream": 0,      "correlation": 549, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 549, "pid": 2, "tid": 0, "ts": 7429564830372.878,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564830379.252, "dur": 4.089,
    "args": {
      "External id": 1558,      "device": 2, "stream": 0,      "correlation": 552, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 552, "pid": 2, "tid": 0, "ts": 7429564830379.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564830430.169, "dur": 83.632,
    "args": {
      "External id": 1566,      "device": 2, "stream": 0,      "correlation": 556, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 556, "pid": 2, "tid": 0, "ts": 7429564830430.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564830513.802, "dur": 4.770,
    "args": {
      "External id": 1566,      "device": 2, "stream": 0,      "correlation": 557, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 557, "pid": 2, "tid": 0, "ts": 7429564830513.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564830518.573, "dur": 6.053,
    "args": {
      "External id": 1569,      "device": 2, "stream": 0,      "correlation": 560, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 560, "pid": 2, "tid": 0, "ts": 7429564830518.573,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564830524.627, "dur": 86.838,
    "args": {
      "External id": 1577,      "device": 2, "stream": 0,      "correlation": 564, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 564, "pid": 2, "tid": 0, "ts": 7429564830524.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564830611.466, "dur": 5.853,
    "args": {
      "External id": 1577,      "device": 2, "stream": 0,      "correlation": 565, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 565, "pid": 2, "tid": 0, "ts": 7429564830611.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564830617.320, "dur": 4.649,
    "args": {
      "External id": 1580,      "device": 2, "stream": 0,      "correlation": 566, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 566, "pid": 2, "tid": 0, "ts": 7429564830617.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564830621.970, "dur": 5.692,
    "args": {
      "External id": 1582,      "device": 2, "stream": 0,      "correlation": 567, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 567, "pid": 2, "tid": 0, "ts": 7429564830621.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564830627.663, "dur": 7.176,
    "args": {
      "External id": 1585,      "device": 2, "stream": 0,      "correlation": 568, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 568, "pid": 2, "tid": 0, "ts": 7429564830627.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564830634.840, "dur": 4.730,
    "args": {
      "External id": 1586,      "device": 2, "stream": 0,      "correlation": 569, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 569, "pid": 2, "tid": 0, "ts": 7429564830634.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564830639.571, "dur": 4.569,
    "args": {
      "External id": 1587,      "device": 2, "stream": 0,      "correlation": 570, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 570, "pid": 2, "tid": 0, "ts": 7429564830639.571,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830644.141, "dur": 4.730,
    "args": {
      "External id": 1588,      "device": 2, "stream": 0,      "correlation": 571, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 571, "pid": 2, "tid": 0, "ts": 7429564830644.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830648.872, "dur": 4.088,
    "args": {
      "External id": 1590,      "device": 2, "stream": 0,      "correlation": 572, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 572, "pid": 2, "tid": 0, "ts": 7429564830648.872,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564830693.575, "dur": 67.233,
    "args": {
      "External id": 1598,      "device": 2, "stream": 0,      "correlation": 576, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 576, "pid": 2, "tid": 0, "ts": 7429564830693.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564830760.809, "dur": 15.555,
    "args": {
      "External id": 1611,      "device": 2, "stream": 0,      "correlation": 580, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 580, "pid": 2, "tid": 0, "ts": 7429564830760.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564830776.365, "dur": 4.129,
    "args": {
      "External id": 1611,      "device": 2, "stream": 0,      "correlation": 581, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 581, "pid": 2, "tid": 0, "ts": 7429564830776.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564830802.666, "dur": 17.519,
    "args": {
      "External id": 1624,      "device": 2, "stream": 0,      "correlation": 585, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 585, "pid": 2, "tid": 0, "ts": 7429564830802.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564830820.186, "dur": 3.928,
    "args": {
      "External id": 1624,      "device": 2, "stream": 0,      "correlation": 586, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 586, "pid": 2, "tid": 0, "ts": 7429564830820.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830833.737, "dur": 3.567,
    "args": {
      "External id": 1634,      "device": 2, "stream": 0,      "correlation": 587, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 587, "pid": 2, "tid": 0, "ts": 7429564830833.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830861.962, "dur": 3.407,
    "args": {
      "External id": 1639,      "device": 2, "stream": 0,      "correlation": 588, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 588, "pid": 2, "tid": 0, "ts": 7429564830861.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564830884.935, "dur": 5.531,
    "args": {
      "External id": 1640,      "device": 2, "stream": 0,      "correlation": 589, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 589, "pid": 2, "tid": 0, "ts": 7429564830884.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830895.760, "dur": 3.527,
    "args": {
      "External id": 1641,      "device": 2, "stream": 0,      "correlation": 590, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 590, "pid": 2, "tid": 0, "ts": 7429564830895.760,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830905.702, "dur": 3.568,
    "args": {
      "External id": 1642,      "device": 2, "stream": 0,      "correlation": 591, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 591, "pid": 2, "tid": 0, "ts": 7429564830905.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830916.487, "dur": 3.207,
    "args": {
      "External id": 1643,      "device": 2, "stream": 0,      "correlation": 592, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 592, "pid": 2, "tid": 0, "ts": 7429564830916.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830932.123, "dur": 3.527,
    "args": {
      "External id": 1648,      "device": 2, "stream": 0,      "correlation": 593, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 593, "pid": 2, "tid": 0, "ts": 7429564830932.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564830944.391, "dur": 6.294,
    "args": {
      "External id": 1649,      "device": 2, "stream": 0,      "correlation": 594, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 594, "pid": 2, "tid": 0, "ts": 7429564830944.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830955.136, "dur": 2.164,
    "args": {
      "External id": 1650,      "device": 2, "stream": 0,      "correlation": 595, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 595, "pid": 2, "tid": 0, "ts": 7429564830955.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564830964.558, "dur": 3.767,
    "args": {
      "External id": 1651,      "device": 2, "stream": 0,      "correlation": 596, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 596, "pid": 2, "tid": 0, "ts": 7429564830964.558,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564831030.068, "dur": 3.968,
    "args": {
      "External id": 1668,      "device": 2, "stream": 0,      "correlation": 597, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 597, "pid": 2, "tid": 0, "ts": 7429564831030.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564831063.986, "dur": 3.848,
    "args": {
      "External id": 1686,      "device": 2, "stream": 0,      "correlation": 598, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 598, "pid": 2, "tid": 0, "ts": 7429564831063.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564831115.064, "dur": 6.053,
    "args": {
      "External id": 1699,      "device": 2, "stream": 0,      "correlation": 602, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 602, "pid": 2, "tid": 0, "ts": 7429564831115.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831131.662, "dur": 4.329,
    "args": {
      "External id": 1702,      "device": 2, "stream": 0,      "correlation": 603, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 603, "pid": 2, "tid": 0, "ts": 7429564831131.662,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564831150.425, "dur": 3.287,
    "args": {
      "External id": 1709,      "device": 2, "stream": 0,      "correlation": 604, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 604, "pid": 2, "tid": 0, "ts": 7429564831150.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564831167.985, "dur": 3.527,
    "args": {
      "External id": 1712,      "device": 2, "stream": 0,      "correlation": 605, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 605, "pid": 2, "tid": 0, "ts": 7429564831167.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564831223.954, "dur": 5.933,
    "args": {
      "External id": 1724,      "device": 2, "stream": 0,      "correlation": 609, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 609, "pid": 2, "tid": 0, "ts": 7429564831223.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564831244.762, "dur": 4.128,
    "args": {
      "External id": 1733,      "device": 2, "stream": 0,      "correlation": 610, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 610, "pid": 2, "tid": 0, "ts": 7429564831244.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564831298.926, "dur": 68.677,
    "args": {
      "External id": 1743,      "device": 2, "stream": 0,      "correlation": 614, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 614, "pid": 2, "tid": 0, "ts": 7429564831298.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831367.604, "dur": 4.650,
    "args": {
      "External id": 1746,      "device": 2, "stream": 0,      "correlation": 615, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 615, "pid": 2, "tid": 0, "ts": 7429564831367.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831372.255, "dur": 4.890,
    "args": {
      "External id": 1748,      "device": 2, "stream": 0,      "correlation": 616, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 616, "pid": 2, "tid": 0, "ts": 7429564831372.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564831377.146, "dur": 7.296,
    "args": {
      "External id": 1751,      "device": 2, "stream": 0,      "correlation": 617, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 617, "pid": 2, "tid": 0, "ts": 7429564831377.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831384.443, "dur": 4.569,
    "args": {
      "External id": 1752,      "device": 2, "stream": 0,      "correlation": 618, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 618, "pid": 2, "tid": 0, "ts": 7429564831384.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831389.013, "dur": 4.650,
    "args": {
      "External id": 1753,      "device": 2, "stream": 0,      "correlation": 619, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 619, "pid": 2, "tid": 0, "ts": 7429564831389.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564831393.664, "dur": 4.369,
    "args": {
      "External id": 1754,      "device": 2, "stream": 0,      "correlation": 620, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 620, "pid": 2, "tid": 0, "ts": 7429564831393.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564831402.605, "dur": 2.565,
    "args": {
      "External id": 1756,      "device": 2, "stream": 0,      "correlation": 621, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 621, "pid": 2, "tid": 0, "ts": 7429564831402.605,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564831530.178, "dur": 83.551,
    "args": {
      "External id": 1764,      "device": 2, "stream": 0,      "correlation": 627, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 627, "pid": 2, "tid": 0, "ts": 7429564831530.178,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564831613.730, "dur": 4.770,
    "args": {
      "External id": 1764,      "device": 2, "stream": 0,      "correlation": 628, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 628, "pid": 2, "tid": 0, "ts": 7429564831613.730,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831618.501, "dur": 4.209,
    "args": {
      "External id": 1767,      "device": 2, "stream": 0,      "correlation": 629, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 629, "pid": 2, "tid": 0, "ts": 7429564831618.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564831643.719, "dur": 81.707,
    "args": {
      "External id": 1775,      "device": 2, "stream": 0,      "correlation": 635, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 635, "pid": 2, "tid": 0, "ts": 7429564831643.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564831725.427, "dur": 5.090,
    "args": {
      "External id": 1775,      "device": 2, "stream": 0,      "correlation": 636, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 636, "pid": 2, "tid": 0, "ts": 7429564831725.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831730.518, "dur": 4.971,
    "args": {
      "External id": 1778,      "device": 2, "stream": 0,      "correlation": 637, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 637, "pid": 2, "tid": 0, "ts": 7429564831730.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564831735.490, "dur": 87.480,
    "args": {
      "External id": 1786,      "device": 2, "stream": 0,      "correlation": 641, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 641, "pid": 2, "tid": 0, "ts": 7429564831735.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564831822.971, "dur": 4.650,
    "args": {
      "External id": 1786,      "device": 2, "stream": 0,      "correlation": 642, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 642, "pid": 2, "tid": 0, "ts": 7429564831822.971,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831827.622, "dur": 4.689,
    "args": {
      "External id": 1789,      "device": 2, "stream": 0,      "correlation": 643, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 643, "pid": 2, "tid": 0, "ts": 7429564831827.622,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831832.312, "dur": 4.730,
    "args": {
      "External id": 1791,      "device": 2, "stream": 0,      "correlation": 644, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 644, "pid": 2, "tid": 0, "ts": 7429564831832.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564831837.043, "dur": 7.176,
    "args": {
      "External id": 1794,      "device": 2, "stream": 0,      "correlation": 645, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 645, "pid": 2, "tid": 0, "ts": 7429564831837.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831844.220, "dur": 4.489,
    "args": {
      "External id": 1795,      "device": 2, "stream": 0,      "correlation": 646, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 646, "pid": 2, "tid": 0, "ts": 7429564831844.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564831848.710, "dur": 4.529,
    "args": {
      "External id": 1796,      "device": 2, "stream": 0,      "correlation": 647, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 647, "pid": 2, "tid": 0, "ts": 7429564831848.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564831853.240, "dur": 4.690,
    "args": {
      "External id": 1797,      "device": 2, "stream": 0,      "correlation": 648, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 648, "pid": 2, "tid": 0, "ts": 7429564831853.240,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564831857.931, "dur": 4.129,
    "args": {
      "External id": 1799,      "device": 2, "stream": 0,      "correlation": 649, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 649, "pid": 2, "tid": 0, "ts": 7429564831857.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564831874.209, "dur": 67.754,
    "args": {
      "External id": 1807,      "device": 2, "stream": 0,      "correlation": 653, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 653, "pid": 2, "tid": 0, "ts": 7429564831874.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564831941.964, "dur": 16.157,
    "args": {
      "External id": 1820,      "device": 2, "stream": 0,      "correlation": 657, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 657, "pid": 2, "tid": 0, "ts": 7429564831941.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564831958.122, "dur": 3.928,
    "args": {
      "External id": 1820,      "device": 2, "stream": 0,      "correlation": 658, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 658, "pid": 2, "tid": 0, "ts": 7429564831958.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564831981.375, "dur": 16.316,
    "args": {
      "External id": 1833,      "device": 2, "stream": 0,      "correlation": 662, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 662, "pid": 2, "tid": 0, "ts": 7429564831981.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564831997.692, "dur": 3.808,
    "args": {
      "External id": 1833,      "device": 2, "stream": 0,      "correlation": 663, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 663, "pid": 2, "tid": 0, "ts": 7429564831997.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832054.022, "dur": 3.286,
    "args": {
      "External id": 1843,      "device": 2, "stream": 0,      "correlation": 666, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 666, "pid": 2, "tid": 0, "ts": 7429564832054.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832086.737, "dur": 2.445,
    "args": {
      "External id": 1848,      "device": 2, "stream": 0,      "correlation": 667, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 667, "pid": 2, "tid": 0, "ts": 7429564832086.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564832107.024, "dur": 6.614,
    "args": {
      "External id": 1849,      "device": 2, "stream": 0,      "correlation": 668, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 668, "pid": 2, "tid": 0, "ts": 7429564832107.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832120.134, "dur": 3.407,
    "args": {
      "External id": 1850,      "device": 2, "stream": 0,      "correlation": 669, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 669, "pid": 2, "tid": 0, "ts": 7429564832120.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832131.119, "dur": 3.367,
    "args": {
      "External id": 1851,      "device": 2, "stream": 0,      "correlation": 670, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 670, "pid": 2, "tid": 0, "ts": 7429564832131.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832139.779, "dur": 3.287,
    "args": {
      "External id": 1852,      "device": 2, "stream": 0,      "correlation": 671, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 671, "pid": 2, "tid": 0, "ts": 7429564832139.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832158.863, "dur": 2.926,
    "args": {
      "External id": 1857,      "device": 2, "stream": 0,      "correlation": 672, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 672, "pid": 2, "tid": 0, "ts": 7429564832158.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564832172.053, "dur": 5.773,
    "args": {
      "External id": 1858,      "device": 2, "stream": 0,      "correlation": 673, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 673, "pid": 2, "tid": 0, "ts": 7429564832172.053,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832183.680, "dur": 2.525,
    "args": {
      "External id": 1859,      "device": 2, "stream": 0,      "correlation": 674, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 674, "pid": 2, "tid": 0, "ts": 7429564832183.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832190.656, "dur": 3.447,
    "args": {
      "External id": 1860,      "device": 2, "stream": 0,      "correlation": 675, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 675, "pid": 2, "tid": 0, "ts": 7429564832190.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832256.688, "dur": 3.086,
    "args": {
      "External id": 1877,      "device": 2, "stream": 0,      "correlation": 676, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 676, "pid": 2, "tid": 0, "ts": 7429564832256.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832289.403, "dur": 3.687,
    "args": {
      "External id": 1895,      "device": 2, "stream": 0,      "correlation": 677, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 677, "pid": 2, "tid": 0, "ts": 7429564832289.403,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564832340.841, "dur": 5.853,
    "args": {
      "External id": 1908,      "device": 2, "stream": 0,      "correlation": 681, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 681, "pid": 2, "tid": 0, "ts": 7429564832340.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564832358.522, "dur": 3.327,
    "args": {
      "External id": 1911,      "device": 2, "stream": 0,      "correlation": 682, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 682, "pid": 2, "tid": 0, "ts": 7429564832358.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832376.884, "dur": 2.645,
    "args": {
      "External id": 1918,      "device": 2, "stream": 0,      "correlation": 683, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 683, "pid": 2, "tid": 0, "ts": 7429564832376.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564832394.324, "dur": 3.527,
    "args": {
      "External id": 1921,      "device": 2, "stream": 0,      "correlation": 684, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 684, "pid": 2, "tid": 0, "ts": 7429564832394.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564832450.373, "dur": 5.411,
    "args": {
      "External id": 1933,      "device": 2, "stream": 0,      "correlation": 688, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 688, "pid": 2, "tid": 0, "ts": 7429564832450.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832469.858, "dur": 3.407,
    "args": {
      "External id": 1942,      "device": 2, "stream": 0,      "correlation": 689, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 689, "pid": 2, "tid": 0, "ts": 7429564832469.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564832523.902, "dur": 67.995,
    "args": {
      "External id": 1952,      "device": 2, "stream": 0,      "correlation": 693, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 693, "pid": 2, "tid": 0, "ts": 7429564832523.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564832591.898, "dur": 4.730,
    "args": {
      "External id": 1955,      "device": 2, "stream": 0,      "correlation": 694, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 694, "pid": 2, "tid": 0, "ts": 7429564832591.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564832596.629, "dur": 4.971,
    "args": {
      "External id": 1957,      "device": 2, "stream": 0,      "correlation": 695, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 695, "pid": 2, "tid": 0, "ts": 7429564832596.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564832601.601, "dur": 6.935,
    "args": {
      "External id": 1960,      "device": 2, "stream": 0,      "correlation": 696, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 696, "pid": 2, "tid": 0, "ts": 7429564832601.601,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564832608.537, "dur": 4.569,
    "args": {
      "External id": 1961,      "device": 2, "stream": 0,      "correlation": 697, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 697, "pid": 2, "tid": 0, "ts": 7429564832608.537,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564832613.107, "dur": 5.852,
    "args": {
      "External id": 1962,      "device": 2, "stream": 0,      "correlation": 698, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 698, "pid": 2, "tid": 0, "ts": 7429564832613.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832618.960, "dur": 4.249,
    "args": {
      "External id": 1963,      "device": 2, "stream": 0,      "correlation": 699, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 699, "pid": 2, "tid": 0, "ts": 7429564832618.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564832629.465, "dur": 4.128,
    "args": {
      "External id": 1965,      "device": 2, "stream": 0,      "correlation": 700, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 700, "pid": 2, "tid": 0, "ts": 7429564832629.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564832684.391, "dur": 82.789,
    "args": {
      "External id": 1973,      "device": 2, "stream": 0,      "correlation": 704, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 704, "pid": 2, "tid": 0, "ts": 7429564832684.391,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564832767.181, "dur": 4.770,
    "args": {
      "External id": 1973,      "device": 2, "stream": 0,      "correlation": 705, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 705, "pid": 2, "tid": 0, "ts": 7429564832767.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564832771.952, "dur": 4.089,
    "args": {
      "External id": 1976,      "device": 2, "stream": 0,      "correlation": 708, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 708, "pid": 2, "tid": 0, "ts": 7429564832771.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564832811.122, "dur": 81.627,
    "args": {
      "External id": 1984,      "device": 2, "stream": 0,      "correlation": 712, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 712, "pid": 2, "tid": 0, "ts": 7429564832811.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564832892.750, "dur": 4.730,
    "args": {
      "External id": 1984,      "device": 2, "stream": 0,      "correlation": 713, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 713, "pid": 2, "tid": 0, "ts": 7429564832892.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564832897.481, "dur": 5.211,
    "args": {
      "External id": 1987,      "device": 2, "stream": 0,      "correlation": 716, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 716, "pid": 2, "tid": 0, "ts": 7429564832897.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564832902.693, "dur": 86.518,
    "args": {
      "External id": 1995,      "device": 2, "stream": 0,      "correlation": 720, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 720, "pid": 2, "tid": 0, "ts": 7429564832902.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564832989.212, "dur": 5.371,
    "args": {
      "External id": 1995,      "device": 2, "stream": 0,      "correlation": 721, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 721, "pid": 2, "tid": 0, "ts": 7429564832989.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564832994.584, "dur": 4.690,
    "args": {
      "External id": 1998,      "device": 2, "stream": 0,      "correlation": 722, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 722, "pid": 2, "tid": 0, "ts": 7429564832994.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564832999.275, "dur": 5.491,
    "args": {
      "External id": 2000,      "device": 2, "stream": 0,      "correlation": 723, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 723, "pid": 2, "tid": 0, "ts": 7429564832999.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564833004.767, "dur": 7.096,
    "args": {
      "External id": 2003,      "device": 2, "stream": 0,      "correlation": 724, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 724, "pid": 2, "tid": 0, "ts": 7429564833004.767,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564833011.864, "dur": 4.409,
    "args": {
      "External id": 2004,      "device": 2, "stream": 0,      "correlation": 725, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 725, "pid": 2, "tid": 0, "ts": 7429564833011.864,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564833016.274, "dur": 4.529,
    "args": {
      "External id": 2005,      "device": 2, "stream": 0,      "correlation": 726, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 726, "pid": 2, "tid": 0, "ts": 7429564833016.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833020.804, "dur": 4.810,
    "args": {
      "External id": 2006,      "device": 2, "stream": 0,      "correlation": 727, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 727, "pid": 2, "tid": 0, "ts": 7429564833020.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833025.615, "dur": 4.089,
    "args": {
      "External id": 2008,      "device": 2, "stream": 0,      "correlation": 728, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 728, "pid": 2, "tid": 0, "ts": 7429564833025.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564833067.471, "dur": 68.196,
    "args": {
      "External id": 2016,      "device": 2, "stream": 0,      "correlation": 732, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 732, "pid": 2, "tid": 0, "ts": 7429564833067.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564833283.929, "dur": 16.757,
    "args": {
      "External id": 2029,      "device": 2, "stream": 0,      "correlation": 736, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 736, "pid": 2, "tid": 0, "ts": 7429564833283.929,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564833300.687, "dur": 4.209,
    "args": {
      "External id": 2029,      "device": 2, "stream": 0,      "correlation": 737, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 737, "pid": 2, "tid": 0, "ts": 7429564833300.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564833340.659, "dur": 17.159,
    "args": {
      "External id": 2042,      "device": 2, "stream": 0,      "correlation": 741, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 741, "pid": 2, "tid": 0, "ts": 7429564833340.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564833357.819, "dur": 4.048,
    "args": {
      "External id": 2042,      "device": 2, "stream": 0,      "correlation": 742, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 742, "pid": 2, "tid": 0, "ts": 7429564833357.819,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833405.689, "dur": 3.848,
    "args": {
      "External id": 2052,      "device": 2, "stream": 0,      "correlation": 743, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 743, "pid": 2, "tid": 0, "ts": 7429564833405.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833437.201, "dur": 3.006,
    "args": {
      "External id": 2057,      "device": 2, "stream": 0,      "correlation": 744, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 744, "pid": 2, "tid": 0, "ts": 7429564833437.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564833461.176, "dur": 5.773,
    "args": {
      "External id": 2058,      "device": 2, "stream": 0,      "correlation": 745, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 745, "pid": 2, "tid": 0, "ts": 7429564833461.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833472.883, "dur": 3.046,
    "args": {
      "External id": 2059,      "device": 2, "stream": 0,      "correlation": 746, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 746, "pid": 2, "tid": 0, "ts": 7429564833472.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833484.751, "dur": 2.645,
    "args": {
      "External id": 2060,      "device": 2, "stream": 0,      "correlation": 747, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 747, "pid": 2, "tid": 0, "ts": 7429564833484.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833494.252, "dur": 2.605,
    "args": {
      "External id": 2061,      "device": 2, "stream": 0,      "correlation": 748, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 748, "pid": 2, "tid": 0, "ts": 7429564833494.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833510.570, "dur": 3.487,
    "args": {
      "External id": 2066,      "device": 2, "stream": 0,      "correlation": 749, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 749, "pid": 2, "tid": 0, "ts": 7429564833510.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564833522.237, "dur": 5.732,
    "args": {
      "External id": 2067,      "device": 2, "stream": 0,      "correlation": 750, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 750, "pid": 2, "tid": 0, "ts": 7429564833522.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833533.743, "dur": 2.485,
    "args": {
      "External id": 2068,      "device": 2, "stream": 0,      "correlation": 751, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 751, "pid": 2, "tid": 0, "ts": 7429564833533.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833542.563, "dur": 3.327,
    "args": {
      "External id": 2069,      "device": 2, "stream": 0,      "correlation": 752, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 752, "pid": 2, "tid": 0, "ts": 7429564833542.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833611.642, "dur": 2.966,
    "args": {
      "External id": 2086,      "device": 2, "stream": 0,      "correlation": 753, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 753, "pid": 2, "tid": 0, "ts": 7429564833611.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833646.122, "dur": 3.086,
    "args": {
      "External id": 2104,      "device": 2, "stream": 0,      "correlation": 754, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 754, "pid": 2, "tid": 0, "ts": 7429564833646.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564833697.279, "dur": 5.251,
    "args": {
      "External id": 2117,      "device": 2, "stream": 0,      "correlation": 758, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 758, "pid": 2, "tid": 0, "ts": 7429564833697.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564833714.439, "dur": 3.406,
    "args": {
      "External id": 2120,      "device": 2, "stream": 0,      "correlation": 759, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 759, "pid": 2, "tid": 0, "ts": 7429564833714.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833732.239, "dur": 2.846,
    "args": {
      "External id": 2127,      "device": 2, "stream": 0,      "correlation": 760, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 760, "pid": 2, "tid": 0, "ts": 7429564833732.239,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564833750.642, "dur": 3.487,
    "args": {
      "External id": 2130,      "device": 2, "stream": 0,      "correlation": 761, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 761, "pid": 2, "tid": 0, "ts": 7429564833750.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564833806.009, "dur": 5.492,
    "args": {
      "External id": 2142,      "device": 2, "stream": 0,      "correlation": 765, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 765, "pid": 2, "tid": 0, "ts": 7429564833806.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564833826.737, "dur": 3.406,
    "args": {
      "External id": 2151,      "device": 2, "stream": 0,      "correlation": 766, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 766, "pid": 2, "tid": 0, "ts": 7429564833826.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564833881.703, "dur": 66.953,
    "args": {
      "External id": 2161,      "device": 2, "stream": 0,      "correlation": 770, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 770, "pid": 2, "tid": 0, "ts": 7429564833881.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564833965.415, "dur": 2.605,
    "args": {
      "External id": 2164,      "device": 2, "stream": 0,      "correlation": 773, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 773, "pid": 2, "tid": 0, "ts": 7429564833965.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564834008.274, "dur": 2.605,
    "args": {
      "External id": 2166,      "device": 2, "stream": 0,      "correlation": 774, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 774, "pid": 2, "tid": 0, "ts": 7429564834008.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564834026.315, "dur": 7.577,
    "args": {
      "External id": 2169,      "device": 2, "stream": 0,      "correlation": 775, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 775, "pid": 2, "tid": 0, "ts": 7429564834026.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564834039.385, "dur": 2.966,
    "args": {
      "External id": 2170,      "device": 2, "stream": 0,      "correlation": 776, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 776, "pid": 2, "tid": 0, "ts": 7429564834039.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564834050.691, "dur": 2.806,
    "args": {
      "External id": 2171,      "device": 2, "stream": 0,      "correlation": 777, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 777, "pid": 2, "tid": 0, "ts": 7429564834050.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834062.639, "dur": 2.805,
    "args": {
      "External id": 2172,      "device": 2, "stream": 0,      "correlation": 778, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 778, "pid": 2, "tid": 0, "ts": 7429564834062.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834072.902, "dur": 3.167,
    "args": {
      "External id": 2174,      "device": 2, "stream": 0,      "correlation": 779, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 779, "pid": 2, "tid": 0, "ts": 7429564834072.902,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564834166.478, "dur": 82.147,
    "args": {
      "External id": 2182,      "device": 2, "stream": 0,      "correlation": 783, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 783, "pid": 2, "tid": 0, "ts": 7429564834166.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564834248.626, "dur": 4.570,
    "args": {
      "External id": 2182,      "device": 2, "stream": 0,      "correlation": 784, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 784, "pid": 2, "tid": 0, "ts": 7429564834248.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564834253.197, "dur": 4.048,
    "args": {
      "External id": 2185,      "device": 2, "stream": 0,      "correlation": 785, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 785, "pid": 2, "tid": 0, "ts": 7429564834253.197,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564834294.091, "dur": 85.395,
    "args": {
      "External id": 2193,      "device": 2, "stream": 0,      "correlation": 791, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 791, "pid": 2, "tid": 0, "ts": 7429564834294.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564834379.487, "dur": 4.890,
    "args": {
      "External id": 2193,      "device": 2, "stream": 0,      "correlation": 792, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 792, "pid": 2, "tid": 0, "ts": 7429564834379.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564834384.378, "dur": 5.011,
    "args": {
      "External id": 2196,      "device": 2, "stream": 0,      "correlation": 793, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 793, "pid": 2, "tid": 0, "ts": 7429564834384.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564834389.390, "dur": 89.445,
    "args": {
      "External id": 2204,      "device": 2, "stream": 0,      "correlation": 797, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 797, "pid": 2, "tid": 0, "ts": 7429564834389.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564834478.836, "dur": 4.649,
    "args": {
      "External id": 2204,      "device": 2, "stream": 0,      "correlation": 798, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 798, "pid": 2, "tid": 0, "ts": 7429564834478.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564834483.486, "dur": 4.890,
    "args": {
      "External id": 2207,      "device": 2, "stream": 0,      "correlation": 799, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 799, "pid": 2, "tid": 0, "ts": 7429564834483.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564834488.377, "dur": 4.690,
    "args": {
      "External id": 2209,      "device": 2, "stream": 0,      "correlation": 800, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 800, "pid": 2, "tid": 0, "ts": 7429564834488.377,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564834493.068, "dur": 7.136,
    "args": {
      "External id": 2212,      "device": 2, "stream": 0,      "correlation": 801, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 801, "pid": 2, "tid": 0, "ts": 7429564834493.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564834500.205, "dur": 4.649,
    "args": {
      "External id": 2213,      "device": 2, "stream": 0,      "correlation": 802, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 802, "pid": 2, "tid": 0, "ts": 7429564834500.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564834504.855, "dur": 4.570,
    "args": {
      "External id": 2214,      "device": 2, "stream": 0,      "correlation": 803, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 803, "pid": 2, "tid": 0, "ts": 7429564834504.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834509.426, "dur": 6.293,
    "args": {
      "External id": 2215,      "device": 2, "stream": 0,      "correlation": 804, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 804, "pid": 2, "tid": 0, "ts": 7429564834509.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834515.720, "dur": 4.129,
    "args": {
      "External id": 2217,      "device": 2, "stream": 0,      "correlation": 807, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 807, "pid": 2, "tid": 0, "ts": 7429564834515.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564834576.139, "dur": 68.236,
    "args": {
      "External id": 2225,      "device": 2, "stream": 0,      "correlation": 811, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 811, "pid": 2, "tid": 0, "ts": 7429564834576.139,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564834644.376, "dur": 15.876,
    "args": {
      "External id": 2238,      "device": 2, "stream": 0,      "correlation": 815, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 815, "pid": 2, "tid": 0, "ts": 7429564834644.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564834660.253, "dur": 3.968,
    "args": {
      "External id": 2238,      "device": 2, "stream": 0,      "correlation": 816, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 816, "pid": 2, "tid": 0, "ts": 7429564834660.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564834687.355, "dur": 16.717,
    "args": {
      "External id": 2251,      "device": 2, "stream": 0,      "correlation": 820, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 820, "pid": 2, "tid": 0, "ts": 7429564834687.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564834704.073, "dur": 3.968,
    "args": {
      "External id": 2251,      "device": 2, "stream": 0,      "correlation": 821, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 821, "pid": 2, "tid": 0, "ts": 7429564834704.073,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834719.990, "dur": 3.687,
    "args": {
      "External id": 2261,      "device": 2, "stream": 0,      "correlation": 822, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 822, "pid": 2, "tid": 0, "ts": 7429564834719.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834749.177, "dur": 3.808,
    "args": {
      "External id": 2266,      "device": 2, "stream": 0,      "correlation": 823, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 823, "pid": 2, "tid": 0, "ts": 7429564834749.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564834771.749, "dur": 6.734,
    "args": {
      "External id": 2267,      "device": 2, "stream": 0,      "correlation": 824, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 824, "pid": 2, "tid": 0, "ts": 7429564834771.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834784.057, "dur": 3.567,
    "args": {
      "External id": 2268,      "device": 2, "stream": 0,      "correlation": 825, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 825, "pid": 2, "tid": 0, "ts": 7429564834784.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834794.682, "dur": 3.647,
    "args": {
      "External id": 2269,      "device": 2, "stream": 0,      "correlation": 826, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 826, "pid": 2, "tid": 0, "ts": 7429564834794.682,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834806.308, "dur": 3.327,
    "args": {
      "External id": 2270,      "device": 2, "stream": 0,      "correlation": 827, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 827, "pid": 2, "tid": 0, "ts": 7429564834806.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834824.029, "dur": 3.768,
    "args": {
      "External id": 2275,      "device": 2, "stream": 0,      "correlation": 828, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 828, "pid": 2, "tid": 0, "ts": 7429564834824.029,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564834834.894, "dur": 6.173,
    "args": {
      "External id": 2276,      "device": 2, "stream": 0,      "correlation": 829, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 829, "pid": 2, "tid": 0, "ts": 7429564834834.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834845.438, "dur": 3.608,
    "args": {
      "External id": 2277,      "device": 2, "stream": 0,      "correlation": 830, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 830, "pid": 2, "tid": 0, "ts": 7429564834845.438,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834853.216, "dur": 3.848,
    "args": {
      "External id": 2278,      "device": 2, "stream": 0,      "correlation": 831, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 831, "pid": 2, "tid": 0, "ts": 7429564834853.216,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834921.212, "dur": 4.129,
    "args": {
      "External id": 2295,      "device": 2, "stream": 0,      "correlation": 832, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 832, "pid": 2, "tid": 0, "ts": 7429564834921.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564834955.531, "dur": 3.648,
    "args": {
      "External id": 2313,      "device": 2, "stream": 0,      "correlation": 833, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 833, "pid": 2, "tid": 0, "ts": 7429564834955.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564835010.458, "dur": 5.932,
    "args": {
      "External id": 2326,      "device": 2, "stream": 0,      "correlation": 837, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 837, "pid": 2, "tid": 0, "ts": 7429564835010.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835027.457, "dur": 4.329,
    "args": {
      "External id": 2329,      "device": 2, "stream": 0,      "correlation": 838, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 838, "pid": 2, "tid": 0, "ts": 7429564835027.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835045.217, "dur": 3.447,
    "args": {
      "External id": 2336,      "device": 2, "stream": 0,      "correlation": 839, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 839, "pid": 2, "tid": 0, "ts": 7429564835045.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564835063.219, "dur": 3.447,
    "args": {
      "External id": 2339,      "device": 2, "stream": 0,      "correlation": 840, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 840, "pid": 2, "tid": 0, "ts": 7429564835063.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564835117.664, "dur": 5.371,
    "args": {
      "External id": 2351,      "device": 2, "stream": 0,      "correlation": 844, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 844, "pid": 2, "tid": 0, "ts": 7429564835117.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835138.271, "dur": 3.928,
    "args": {
      "External id": 2360,      "device": 2, "stream": 0,      "correlation": 845, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 845, "pid": 2, "tid": 0, "ts": 7429564835138.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564835193.117, "dur": 68.597,
    "args": {
      "External id": 2370,      "device": 2, "stream": 0,      "correlation": 849, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 849, "pid": 2, "tid": 0, "ts": 7429564835193.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835261.715, "dur": 4.730,
    "args": {
      "External id": 2373,      "device": 2, "stream": 0,      "correlation": 850, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 850, "pid": 2, "tid": 0, "ts": 7429564835261.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835266.446, "dur": 4.850,
    "args": {
      "External id": 2375,      "device": 2, "stream": 0,      "correlation": 851, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 851, "pid": 2, "tid": 0, "ts": 7429564835266.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564835271.297, "dur": 7.135,
    "args": {
      "External id": 2378,      "device": 2, "stream": 0,      "correlation": 852, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 852, "pid": 2, "tid": 0, "ts": 7429564835271.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835278.433, "dur": 4.570,
    "args": {
      "External id": 2379,      "device": 2, "stream": 0,      "correlation": 853, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 853, "pid": 2, "tid": 0, "ts": 7429564835278.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835283.004, "dur": 4.570,
    "args": {
      "External id": 2380,      "device": 2, "stream": 0,      "correlation": 854, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 854, "pid": 2, "tid": 0, "ts": 7429564835283.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835287.575, "dur": 4.328,
    "args": {
      "External id": 2381,      "device": 2, "stream": 0,      "correlation": 855, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 855, "pid": 2, "tid": 0, "ts": 7429564835287.575,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835296.194, "dur": 3.568,
    "args": {
      "External id": 2383,      "device": 2, "stream": 0,      "correlation": 856, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 856, "pid": 2, "tid": 0, "ts": 7429564835296.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564835411.459, "dur": 81.386,
    "args": {
      "External id": 2391,      "device": 2, "stream": 0,      "correlation": 862, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 862, "pid": 2, "tid": 0, "ts": 7429564835411.459,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564835492.846, "dur": 4.971,
    "args": {
      "External id": 2391,      "device": 2, "stream": 0,      "correlation": 863, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 863, "pid": 2, "tid": 0, "ts": 7429564835492.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835497.818, "dur": 4.048,
    "args": {
      "External id": 2394,      "device": 2, "stream": 0,      "correlation": 864, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 864, "pid": 2, "tid": 0, "ts": 7429564835497.818,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564835518.185, "dur": 77.978,
    "args": {
      "External id": 2402,      "device": 2, "stream": 0,      "correlation": 870, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 870, "pid": 2, "tid": 0, "ts": 7429564835518.185,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564835596.164, "dur": 5.050,
    "args": {
      "External id": 2402,      "device": 2, "stream": 0,      "correlation": 871, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 871, "pid": 2, "tid": 0, "ts": 7429564835596.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835601.215, "dur": 4.971,
    "args": {
      "External id": 2405,      "device": 2, "stream": 0,      "correlation": 872, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 872, "pid": 2, "tid": 0, "ts": 7429564835601.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564835606.187, "dur": 87.721,
    "args": {
      "External id": 2413,      "device": 2, "stream": 0,      "correlation": 876, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 876, "pid": 2, "tid": 0, "ts": 7429564835606.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564835693.909, "dur": 4.569,
    "args": {
      "External id": 2413,      "device": 2, "stream": 0,      "correlation": 877, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 877, "pid": 2, "tid": 0, "ts": 7429564835693.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835698.479, "dur": 4.570,
    "args": {
      "External id": 2416,      "device": 2, "stream": 0,      "correlation": 878, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 878, "pid": 2, "tid": 0, "ts": 7429564835698.479,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835703.050, "dur": 4.569,
    "args": {
      "External id": 2418,      "device": 2, "stream": 0,      "correlation": 879, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 879, "pid": 2, "tid": 0, "ts": 7429564835703.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564835707.620, "dur": 7.296,
    "args": {
      "External id": 2421,      "device": 2, "stream": 0,      "correlation": 880, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 880, "pid": 2, "tid": 0, "ts": 7429564835707.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835714.917, "dur": 4.449,
    "args": {
      "External id": 2422,      "device": 2, "stream": 0,      "correlation": 881, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 881, "pid": 2, "tid": 0, "ts": 7429564835714.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564835719.367, "dur": 4.529,
    "args": {
      "External id": 2423,      "device": 2, "stream": 0,      "correlation": 882, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 882, "pid": 2, "tid": 0, "ts": 7429564835719.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835723.897, "dur": 4.730,
    "args": {
      "External id": 2424,      "device": 2, "stream": 0,      "correlation": 883, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 883, "pid": 2, "tid": 0, "ts": 7429564835723.897,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835728.628, "dur": 4.209,
    "args": {
      "External id": 2426,      "device": 2, "stream": 0,      "correlation": 884, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 884, "pid": 2, "tid": 0, "ts": 7429564835728.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564835740.175, "dur": 68.156,
    "args": {
      "External id": 2434,      "device": 2, "stream": 0,      "correlation": 888, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 888, "pid": 2, "tid": 0, "ts": 7429564835740.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564835808.332, "dur": 15.755,
    "args": {
      "External id": 2447,      "device": 2, "stream": 0,      "correlation": 892, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 892, "pid": 2, "tid": 0, "ts": 7429564835808.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564835824.088, "dur": 4.168,
    "args": {
      "External id": 2447,      "device": 2, "stream": 0,      "correlation": 893, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 893, "pid": 2, "tid": 0, "ts": 7429564835824.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564835840.766, "dur": 16.677,
    "args": {
      "External id": 2460,      "device": 2, "stream": 0,      "correlation": 897, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 897, "pid": 2, "tid": 0, "ts": 7429564835840.766,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564835857.444, "dur": 3.928,
    "args": {
      "External id": 2460,      "device": 2, "stream": 0,      "correlation": 898, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 898, "pid": 2, "tid": 0, "ts": 7429564835857.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835875.125, "dur": 3.086,
    "args": {
      "External id": 2470,      "device": 2, "stream": 0,      "correlation": 899, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 899, "pid": 2, "tid": 0, "ts": 7429564835875.125,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835903.390, "dur": 2.324,
    "args": {
      "External id": 2475,      "device": 2, "stream": 0,      "correlation": 900, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 900, "pid": 2, "tid": 0, "ts": 7429564835903.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564835963.729, "dur": 6.093,
    "args": {
      "External id": 2476,      "device": 2, "stream": 0,      "correlation": 903, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 903, "pid": 2, "tid": 0, "ts": 7429564835963.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835976.558, "dur": 2.846,
    "args": {
      "External id": 2477,      "device": 2, "stream": 0,      "correlation": 904, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 904, "pid": 2, "tid": 0, "ts": 7429564835976.558,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835985.699, "dur": 3.848,
    "args": {
      "External id": 2478,      "device": 2, "stream": 0,      "correlation": 905, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 905, "pid": 2, "tid": 0, "ts": 7429564835985.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564835995.923, "dur": 3.567,
    "args": {
      "External id": 2479,      "device": 2, "stream": 0,      "correlation": 906, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 906, "pid": 2, "tid": 0, "ts": 7429564835995.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836013.523, "dur": 3.728,
    "args": {
      "External id": 2484,      "device": 2, "stream": 0,      "correlation": 907, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 907, "pid": 2, "tid": 0, "ts": 7429564836013.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564836028.437, "dur": 6.214,
    "args": {
      "External id": 2485,      "device": 2, "stream": 0,      "correlation": 908, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 908, "pid": 2, "tid": 0, "ts": 7429564836028.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836038.340, "dur": 3.287,
    "args": {
      "External id": 2486,      "device": 2, "stream": 0,      "correlation": 909, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 909, "pid": 2, "tid": 0, "ts": 7429564836038.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836046.198, "dur": 3.648,
    "args": {
      "External id": 2487,      "device": 2, "stream": 0,      "correlation": 910, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 910, "pid": 2, "tid": 0, "ts": 7429564836046.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836110.105, "dur": 3.688,
    "args": {
      "External id": 2504,      "device": 2, "stream": 0,      "correlation": 911, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 911, "pid": 2, "tid": 0, "ts": 7429564836110.105,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836144.665, "dur": 4.088,
    "args": {
      "External id": 2522,      "device": 2, "stream": 0,      "correlation": 912, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 912, "pid": 2, "tid": 0, "ts": 7429564836144.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564836197.626, "dur": 5.773,
    "args": {
      "External id": 2535,      "device": 2, "stream": 0,      "correlation": 916, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 916, "pid": 2, "tid": 0, "ts": 7429564836197.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836214.866, "dur": 4.449,
    "args": {
      "External id": 2538,      "device": 2, "stream": 0,      "correlation": 917, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 917, "pid": 2, "tid": 0, "ts": 7429564836214.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836233.429, "dur": 3.647,
    "args": {
      "External id": 2545,      "device": 2, "stream": 0,      "correlation": 918, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 918, "pid": 2, "tid": 0, "ts": 7429564836233.429,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564836251.711, "dur": 3.607,
    "args": {
      "External id": 2548,      "device": 2, "stream": 0,      "correlation": 919, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 919, "pid": 2, "tid": 0, "ts": 7429564836251.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564836340.194, "dur": 5.532,
    "args": {
      "External id": 2560,      "device": 2, "stream": 0,      "correlation": 923, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 923, "pid": 2, "tid": 0, "ts": 7429564836340.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836360.080, "dur": 3.808,
    "args": {
      "External id": 2569,      "device": 2, "stream": 0,      "correlation": 924, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 924, "pid": 2, "tid": 0, "ts": 7429564836360.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564836417.572, "dur": 69.078,
    "args": {
      "External id": 2579,      "device": 2, "stream": 0,      "correlation": 928, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 928, "pid": 2, "tid": 0, "ts": 7429564836417.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836486.651, "dur": 4.810,
    "args": {
      "External id": 2582,      "device": 2, "stream": 0,      "correlation": 929, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 929, "pid": 2, "tid": 0, "ts": 7429564836486.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836491.462, "dur": 4.690,
    "args": {
      "External id": 2584,      "device": 2, "stream": 0,      "correlation": 930, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 930, "pid": 2, "tid": 0, "ts": 7429564836491.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564836496.153, "dur": 6.934,
    "args": {
      "External id": 2587,      "device": 2, "stream": 0,      "correlation": 931, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 931, "pid": 2, "tid": 0, "ts": 7429564836496.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836503.088, "dur": 4.530,
    "args": {
      "External id": 2588,      "device": 2, "stream": 0,      "correlation": 932, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 932, "pid": 2, "tid": 0, "ts": 7429564836503.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836507.619, "dur": 4.730,
    "args": {
      "External id": 2589,      "device": 2, "stream": 0,      "correlation": 933, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 933, "pid": 2, "tid": 0, "ts": 7429564836507.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836512.350, "dur": 4.208,
    "args": {
      "External id": 2590,      "device": 2, "stream": 0,      "correlation": 934, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 934, "pid": 2, "tid": 0, "ts": 7429564836512.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836525.741, "dur": 3.767,
    "args": {
      "External id": 2592,      "device": 2, "stream": 0,      "correlation": 935, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 935, "pid": 2, "tid": 0, "ts": 7429564836525.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564836579.143, "dur": 77.698,
    "args": {
      "External id": 2600,      "device": 2, "stream": 0,      "correlation": 939, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 939, "pid": 2, "tid": 0, "ts": 7429564836579.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564836656.842, "dur": 5.612,
    "args": {
      "External id": 2600,      "device": 2, "stream": 0,      "correlation": 940, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 940, "pid": 2, "tid": 0, "ts": 7429564836656.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836662.455, "dur": 4.128,
    "args": {
      "External id": 2603,      "device": 2, "stream": 0,      "correlation": 943, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 943, "pid": 2, "tid": 0, "ts": 7429564836662.455,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564836707.999, "dur": 83.191,
    "args": {
      "External id": 2611,      "device": 2, "stream": 0,      "correlation": 947, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 947, "pid": 2, "tid": 0, "ts": 7429564836707.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564836791.191, "dur": 4.810,
    "args": {
      "External id": 2611,      "device": 2, "stream": 0,      "correlation": 948, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 948, "pid": 2, "tid": 0, "ts": 7429564836791.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836796.002, "dur": 4.088,
    "args": {
      "External id": 2614,      "device": 2, "stream": 0,      "correlation": 951, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 951, "pid": 2, "tid": 0, "ts": 7429564836796.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564836804.140, "dur": 88.884,
    "args": {
      "External id": 2622,      "device": 2, "stream": 0,      "correlation": 955, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 955, "pid": 2, "tid": 0, "ts": 7429564836804.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564836893.025, "dur": 5.050,
    "args": {
      "External id": 2622,      "device": 2, "stream": 0,      "correlation": 956, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 956, "pid": 2, "tid": 0, "ts": 7429564836893.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836898.076, "dur": 4.650,
    "args": {
      "External id": 2625,      "device": 2, "stream": 0,      "correlation": 957, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 957, "pid": 2, "tid": 0, "ts": 7429564836898.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836902.727, "dur": 4.610,
    "args": {
      "External id": 2627,      "device": 2, "stream": 0,      "correlation": 958, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 958, "pid": 2, "tid": 0, "ts": 7429564836902.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564836907.338, "dur": 7.336,
    "args": {
      "External id": 2630,      "device": 2, "stream": 0,      "correlation": 959, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 959, "pid": 2, "tid": 0, "ts": 7429564836907.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836914.675, "dur": 4.489,
    "args": {
      "External id": 2631,      "device": 2, "stream": 0,      "correlation": 960, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 960, "pid": 2, "tid": 0, "ts": 7429564836914.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564836919.165, "dur": 4.449,
    "args": {
      "External id": 2632,      "device": 2, "stream": 0,      "correlation": 961, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 961, "pid": 2, "tid": 0, "ts": 7429564836919.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836923.615, "dur": 4.730,
    "args": {
      "External id": 2633,      "device": 2, "stream": 0,      "correlation": 962, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 962, "pid": 2, "tid": 0, "ts": 7429564836923.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564836928.346, "dur": 4.169,
    "args": {
      "External id": 2635,      "device": 2, "stream": 0,      "correlation": 963, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 963, "pid": 2, "tid": 0, "ts": 7429564836928.346,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564836972.127, "dur": 69.238,
    "args": {
      "External id": 2643,      "device": 2, "stream": 0,      "correlation": 967, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 967, "pid": 2, "tid": 0, "ts": 7429564836972.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564837041.366, "dur": 15.595,
    "args": {
      "External id": 2656,      "device": 2, "stream": 0,      "correlation": 971, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 971, "pid": 2, "tid": 0, "ts": 7429564837041.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564837056.962, "dur": 3.968,
    "args": {
      "External id": 2656,      "device": 2, "stream": 0,      "correlation": 972, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 972, "pid": 2, "tid": 0, "ts": 7429564837056.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564837081.538, "dur": 16.116,
    "args": {
      "External id": 2669,      "device": 2, "stream": 0,      "correlation": 976, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 976, "pid": 2, "tid": 0, "ts": 7429564837081.538,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564837097.655, "dur": 4.048,
    "args": {
      "External id": 2669,      "device": 2, "stream": 0,      "correlation": 977, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 977, "pid": 2, "tid": 0, "ts": 7429564837097.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837112.008, "dur": 3.407,
    "args": {
      "External id": 2679,      "device": 2, "stream": 0,      "correlation": 978, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 978, "pid": 2, "tid": 0, "ts": 7429564837112.008,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837140.474, "dur": 2.765,
    "args": {
      "External id": 2684,      "device": 2, "stream": 0,      "correlation": 979, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 979, "pid": 2, "tid": 0, "ts": 7429564837140.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564837161.803, "dur": 6.574,
    "args": {
      "External id": 2685,      "device": 2, "stream": 0,      "correlation": 980, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 980, "pid": 2, "tid": 0, "ts": 7429564837161.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837175.193, "dur": 2.405,
    "args": {
      "External id": 2686,      "device": 2, "stream": 0,      "correlation": 981, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 981, "pid": 2, "tid": 0, "ts": 7429564837175.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837186.660, "dur": 2.685,
    "args": {
      "External id": 2687,      "device": 2, "stream": 0,      "correlation": 982, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 982, "pid": 2, "tid": 0, "ts": 7429564837186.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837195.320, "dur": 2.645,
    "args": {
      "External id": 2688,      "device": 2, "stream": 0,      "correlation": 983, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 983, "pid": 2, "tid": 0, "ts": 7429564837195.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837212.038, "dur": 2.966,
    "args": {
      "External id": 2693,      "device": 2, "stream": 0,      "correlation": 984, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 984, "pid": 2, "tid": 0, "ts": 7429564837212.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564837222.542, "dur": 5.251,
    "args": {
      "External id": 2694,      "device": 2, "stream": 0,      "correlation": 985, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 985, "pid": 2, "tid": 0, "ts": 7429564837222.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837232.365, "dur": 3.126,
    "args": {
      "External id": 2695,      "device": 2, "stream": 0,      "correlation": 986, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 986, "pid": 2, "tid": 0, "ts": 7429564837232.365,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837241.746, "dur": 3.207,
    "args": {
      "External id": 2696,      "device": 2, "stream": 0,      "correlation": 987, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 987, "pid": 2, "tid": 0, "ts": 7429564837241.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837305.894, "dur": 3.246,
    "args": {
      "External id": 2713,      "device": 2, "stream": 0,      "correlation": 988, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 988, "pid": 2, "tid": 0, "ts": 7429564837305.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837339.972, "dur": 3.327,
    "args": {
      "External id": 2731,      "device": 2, "stream": 0,      "correlation": 989, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 989, "pid": 2, "tid": 0, "ts": 7429564837339.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564837386.279, "dur": 5.611,
    "args": {
      "External id": 2744,      "device": 2, "stream": 0,      "correlation": 993, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 993, "pid": 2, "tid": 0, "ts": 7429564837386.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564837402.075, "dur": 3.527,
    "args": {
      "External id": 2747,      "device": 2, "stream": 0,      "correlation": 994, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 994, "pid": 2, "tid": 0, "ts": 7429564837402.075,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837420.397, "dur": 3.086,
    "args": {
      "External id": 2754,      "device": 2, "stream": 0,      "correlation": 995, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 995, "pid": 2, "tid": 0, "ts": 7429564837420.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564837437.797, "dur": 2.725,
    "args": {
      "External id": 2757,      "device": 2, "stream": 0,      "correlation": 996, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 996, "pid": 2, "tid": 0, "ts": 7429564837437.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564837495.369, "dur": 4.891,
    "args": {
      "External id": 2769,      "device": 2, "stream": 0,      "correlation": 1000, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1000, "pid": 2, "tid": 0, "ts": 7429564837495.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837514.012, "dur": 3.126,
    "args": {
      "External id": 2778,      "device": 2, "stream": 0,      "correlation": 1001, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1001, "pid": 2, "tid": 0, "ts": 7429564837514.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564837568.337, "dur": 67.715,
    "args": {
      "External id": 2788,      "device": 2, "stream": 0,      "correlation": 1005, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1005, "pid": 2, "tid": 0, "ts": 7429564837568.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564837636.053, "dur": 4.649,
    "args": {
      "External id": 2791,      "device": 2, "stream": 0,      "correlation": 1006, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1006, "pid": 2, "tid": 0, "ts": 7429564837636.053,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564837640.703, "dur": 5.412,
    "args": {
      "External id": 2793,      "device": 2, "stream": 0,      "correlation": 1007, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1007, "pid": 2, "tid": 0, "ts": 7429564837640.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564837646.116, "dur": 7.296,
    "args": {
      "External id": 2796,      "device": 2, "stream": 0,      "correlation": 1008, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1008, "pid": 2, "tid": 0, "ts": 7429564837646.116,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564837653.413, "dur": 4.529,
    "args": {
      "External id": 2797,      "device": 2, "stream": 0,      "correlation": 1009, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1009, "pid": 2, "tid": 0, "ts": 7429564837653.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564837657.943, "dur": 4.850,
    "args": {
      "External id": 2798,      "device": 2, "stream": 0,      "correlation": 1010, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1010, "pid": 2, "tid": 0, "ts": 7429564837657.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837662.794, "dur": 4.249,
    "args": {
      "External id": 2799,      "device": 2, "stream": 0,      "correlation": 1011, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1011, "pid": 2, "tid": 0, "ts": 7429564837662.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564837719.324, "dur": 4.249,
    "args": {
      "External id": 2801,      "device": 2, "stream": 0,      "correlation": 1014, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1014, "pid": 2, "tid": 0, "ts": 7429564837719.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564837778.460, "dur": 87.280,
    "args": {
      "External id": 2809,      "device": 2, "stream": 0,      "correlation": 1018, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1018, "pid": 2, "tid": 0, "ts": 7429564837778.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564837865.741, "dur": 4.770,
    "args": {
      "External id": 2809,      "device": 2, "stream": 0,      "correlation": 1019, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1019, "pid": 2, "tid": 0, "ts": 7429564837865.741,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564837870.512, "dur": 4.208,
    "args": {
      "External id": 2812,      "device": 2, "stream": 0,      "correlation": 1020, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1020, "pid": 2, "tid": 0, "ts": 7429564837870.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564837890.237, "dur": 85.235,
    "args": {
      "External id": 2820,      "device": 2, "stream": 0,      "correlation": 1026, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1026, "pid": 2, "tid": 0, "ts": 7429564837890.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564837975.473, "dur": 4.890,
    "args": {
      "External id": 2820,      "device": 2, "stream": 0,      "correlation": 1027, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1027, "pid": 2, "tid": 0, "ts": 7429564837975.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564837980.364, "dur": 5.011,
    "args": {
      "External id": 2823,      "device": 2, "stream": 0,      "correlation": 1028, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1028, "pid": 2, "tid": 0, "ts": 7429564837980.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564837985.376, "dur": 88.201,
    "args": {
      "External id": 2831,      "device": 2, "stream": 0,      "correlation": 1032, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1032, "pid": 2, "tid": 0, "ts": 7429564837985.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564838073.578, "dur": 4.570,
    "args": {
      "External id": 2831,      "device": 2, "stream": 0,      "correlation": 1033, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1033, "pid": 2, "tid": 0, "ts": 7429564838073.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564838078.149, "dur": 4.810,
    "args": {
      "External id": 2834,      "device": 2, "stream": 0,      "correlation": 1034, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1034, "pid": 2, "tid": 0, "ts": 7429564838078.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564838082.960, "dur": 4.650,
    "args": {
      "External id": 2836,      "device": 2, "stream": 0,      "correlation": 1035, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1035, "pid": 2, "tid": 0, "ts": 7429564838082.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564838087.611, "dur": 7.015,
    "args": {
      "External id": 2839,      "device": 2, "stream": 0,      "correlation": 1036, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1036, "pid": 2, "tid": 0, "ts": 7429564838087.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564838094.627, "dur": 4.489,
    "args": {
      "External id": 2840,      "device": 2, "stream": 0,      "correlation": 1037, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1037, "pid": 2, "tid": 0, "ts": 7429564838094.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564838099.117, "dur": 4.610,
    "args": {
      "External id": 2841,      "device": 2, "stream": 0,      "correlation": 1038, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1038, "pid": 2, "tid": 0, "ts": 7429564838099.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838103.728, "dur": 4.810,
    "args": {
      "External id": 2842,      "device": 2, "stream": 0,      "correlation": 1039, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1039, "pid": 2, "tid": 0, "ts": 7429564838103.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838108.539, "dur": 4.248,
    "args": {
      "External id": 2844,      "device": 2, "stream": 0,      "correlation": 1040, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1040, "pid": 2, "tid": 0, "ts": 7429564838108.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564838171.965, "dur": 67.955,
    "args": {
      "External id": 2852,      "device": 2, "stream": 0,      "correlation": 1046, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1046, "pid": 2, "tid": 0, "ts": 7429564838171.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564838239.921, "dur": 15.955,
    "args": {
      "External id": 2865,      "device": 2, "stream": 0,      "correlation": 1050, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1050, "pid": 2, "tid": 0, "ts": 7429564838239.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564838255.877, "dur": 3.888,
    "args": {
      "External id": 2865,      "device": 2, "stream": 0,      "correlation": 1051, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1051, "pid": 2, "tid": 0, "ts": 7429564838255.877,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564838292.762, "dur": 16.958,
    "args": {
      "External id": 2878,      "device": 2, "stream": 0,      "correlation": 1055, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1055, "pid": 2, "tid": 0, "ts": 7429564838292.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564838309.721, "dur": 4.089,
    "args": {
      "External id": 2878,      "device": 2, "stream": 0,      "correlation": 1056, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1056, "pid": 2, "tid": 0, "ts": 7429564838309.721,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838329.046, "dur": 3.406,
    "args": {
      "External id": 2888,      "device": 2, "stream": 0,      "correlation": 1057, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1057, "pid": 2, "tid": 0, "ts": 7429564838329.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838359.997, "dur": 3.527,
    "args": {
      "External id": 2893,      "device": 2, "stream": 0,      "correlation": 1058, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1058, "pid": 2, "tid": 0, "ts": 7429564838359.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564838383.691, "dur": 6.414,
    "args": {
      "External id": 2894,      "device": 2, "stream": 0,      "correlation": 1059, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1059, "pid": 2, "tid": 0, "ts": 7429564838383.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838394.636, "dur": 3.527,
    "args": {
      "External id": 2895,      "device": 2, "stream": 0,      "correlation": 1060, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1060, "pid": 2, "tid": 0, "ts": 7429564838394.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838407.265, "dur": 4.049,
    "args": {
      "External id": 2896,      "device": 2, "stream": 0,      "correlation": 1061, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1061, "pid": 2, "tid": 0, "ts": 7429564838407.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838417.088, "dur": 3.567,
    "args": {
      "External id": 2897,      "device": 2, "stream": 0,      "correlation": 1062, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1062, "pid": 2, "tid": 0, "ts": 7429564838417.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838435.570, "dur": 3.728,
    "args": {
      "External id": 2902,      "device": 2, "stream": 0,      "correlation": 1063, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1063, "pid": 2, "tid": 0, "ts": 7429564838435.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564838447.037, "dur": 5.852,
    "args": {
      "External id": 2903,      "device": 2, "stream": 0,      "correlation": 1064, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1064, "pid": 2, "tid": 0, "ts": 7429564838447.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838457.461, "dur": 3.807,
    "args": {
      "External id": 2904,      "device": 2, "stream": 0,      "correlation": 1065, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1065, "pid": 2, "tid": 0, "ts": 7429564838457.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838466.802, "dur": 2.966,
    "args": {
      "External id": 2905,      "device": 2, "stream": 0,      "correlation": 1066, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1066, "pid": 2, "tid": 0, "ts": 7429564838466.802,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838534.277, "dur": 3.207,
    "args": {
      "External id": 2922,      "device": 2, "stream": 0,      "correlation": 1067, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1067, "pid": 2, "tid": 0, "ts": 7429564838534.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838569.077, "dur": 3.327,
    "args": {
      "External id": 2940,      "device": 2, "stream": 0,      "correlation": 1068, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1068, "pid": 2, "tid": 0, "ts": 7429564838569.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564838620.195, "dur": 5.171,
    "args": {
      "External id": 2953,      "device": 2, "stream": 0,      "correlation": 1072, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1072, "pid": 2, "tid": 0, "ts": 7429564838620.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564838636.993, "dur": 3.287,
    "args": {
      "External id": 2956,      "device": 2, "stream": 0,      "correlation": 1073, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1073, "pid": 2, "tid": 0, "ts": 7429564838636.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838654.353, "dur": 3.006,
    "args": {
      "External id": 2963,      "device": 2, "stream": 0,      "correlation": 1074, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1074, "pid": 2, "tid": 0, "ts": 7429564838654.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564838671.954, "dur": 3.246,
    "args": {
      "External id": 2966,      "device": 2, "stream": 0,      "correlation": 1075, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1075, "pid": 2, "tid": 0, "ts": 7429564838671.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564838729.125, "dur": 4.650,
    "args": {
      "External id": 2978,      "device": 2, "stream": 0,      "correlation": 1079, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1079, "pid": 2, "tid": 0, "ts": 7429564838729.125,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838749.492, "dur": 3.006,
    "args": {
      "External id": 2987,      "device": 2, "stream": 0,      "correlation": 1080, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1080, "pid": 2, "tid": 0, "ts": 7429564838749.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564838802.935, "dur": 68.196,
    "args": {
      "External id": 2997,      "device": 2, "stream": 0,      "correlation": 1084, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1084, "pid": 2, "tid": 0, "ts": 7429564838802.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564838871.132, "dur": 4.930,
    "args": {
      "External id": 3000,      "device": 2, "stream": 0,      "correlation": 1085, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1085, "pid": 2, "tid": 0, "ts": 7429564838871.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564838876.063, "dur": 4.650,
    "args": {
      "External id": 3002,      "device": 2, "stream": 0,      "correlation": 1086, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1086, "pid": 2, "tid": 0, "ts": 7429564838876.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564838880.714, "dur": 6.774,
    "args": {
      "External id": 3005,      "device": 2, "stream": 0,      "correlation": 1087, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1087, "pid": 2, "tid": 0, "ts": 7429564838880.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564838887.489, "dur": 4.530,
    "args": {
      "External id": 3006,      "device": 2, "stream": 0,      "correlation": 1088, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1088, "pid": 2, "tid": 0, "ts": 7429564838887.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564838892.020, "dur": 4.489,
    "args": {
      "External id": 3007,      "device": 2, "stream": 0,      "correlation": 1089, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1089, "pid": 2, "tid": 0, "ts": 7429564838892.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838896.510, "dur": 4.209,
    "args": {
      "External id": 3008,      "device": 2, "stream": 0,      "correlation": 1090, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1090, "pid": 2, "tid": 0, "ts": 7429564838896.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564838909.259, "dur": 3.367,
    "args": {
      "External id": 3010,      "device": 2, "stream": 0,      "correlation": 1091, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1091, "pid": 2, "tid": 0, "ts": 7429564838909.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564839020.194, "dur": 84.233,
    "args": {
      "External id": 3018,      "device": 2, "stream": 0,      "correlation": 1097, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1097, "pid": 2, "tid": 0, "ts": 7429564839020.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564839104.428, "dur": 4.850,
    "args": {
      "External id": 3018,      "device": 2, "stream": 0,      "correlation": 1098, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1098, "pid": 2, "tid": 0, "ts": 7429564839104.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564839109.279, "dur": 4.048,
    "args": {
      "External id": 3021,      "device": 2, "stream": 0,      "correlation": 1099, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1099, "pid": 2, "tid": 0, "ts": 7429564839109.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564839122.830, "dur": 86.879,
    "args": {
      "External id": 3029,      "device": 2, "stream": 0,      "correlation": 1105, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1105, "pid": 2, "tid": 0, "ts": 7429564839122.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564839209.710, "dur": 4.970,
    "args": {
      "External id": 3029,      "device": 2, "stream": 0,      "correlation": 1106, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1106, "pid": 2, "tid": 0, "ts": 7429564839209.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564839214.681, "dur": 4.891,
    "args": {
      "External id": 3032,      "device": 2, "stream": 0,      "correlation": 1107, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1107, "pid": 2, "tid": 0, "ts": 7429564839214.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564839219.573, "dur": 89.324,
    "args": {
      "External id": 3040,      "device": 2, "stream": 0,      "correlation": 1111, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1111, "pid": 2, "tid": 0, "ts": 7429564839219.573,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564839308.898, "dur": 4.650,
    "args": {
      "External id": 3040,      "device": 2, "stream": 0,      "correlation": 1112, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1112, "pid": 2, "tid": 0, "ts": 7429564839308.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564839313.549, "dur": 4.649,
    "args": {
      "External id": 3043,      "device": 2, "stream": 0,      "correlation": 1113, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1113, "pid": 2, "tid": 0, "ts": 7429564839313.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564839318.199, "dur": 4.810,
    "args": {
      "External id": 3045,      "device": 2, "stream": 0,      "correlation": 1114, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1114, "pid": 2, "tid": 0, "ts": 7429564839318.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564839323.010, "dur": 7.096,
    "args": {
      "External id": 3048,      "device": 2, "stream": 0,      "correlation": 1115, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1115, "pid": 2, "tid": 0, "ts": 7429564839323.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564839330.107, "dur": 4.449,
    "args": {
      "External id": 3049,      "device": 2, "stream": 0,      "correlation": 1116, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1116, "pid": 2, "tid": 0, "ts": 7429564839330.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564839334.557, "dur": 4.690,
    "args": {
      "External id": 3050,      "device": 2, "stream": 0,      "correlation": 1117, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1117, "pid": 2, "tid": 0, "ts": 7429564839334.557,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839339.248, "dur": 5.090,
    "args": {
      "External id": 3051,      "device": 2, "stream": 0,      "correlation": 1118, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1118, "pid": 2, "tid": 0, "ts": 7429564839339.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839344.339, "dur": 4.891,
    "args": {
      "External id": 3053,      "device": 2, "stream": 0,      "correlation": 1119, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1119, "pid": 2, "tid": 0, "ts": 7429564839344.339,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564839349.231, "dur": 66.431,
    "args": {
      "External id": 3061,      "device": 2, "stream": 0,      "correlation": 1123, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1123, "pid": 2, "tid": 0, "ts": 7429564839349.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564839427.049, "dur": 17.319,
    "args": {
      "External id": 3074,      "device": 2, "stream": 0,      "correlation": 1127, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1127, "pid": 2, "tid": 0, "ts": 7429564839427.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564839444.369, "dur": 4.008,
    "args": {
      "External id": 3074,      "device": 2, "stream": 0,      "correlation": 1128, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1128, "pid": 2, "tid": 0, "ts": 7429564839444.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564839476.964, "dur": 17.399,
    "args": {
      "External id": 3087,      "device": 2, "stream": 0,      "correlation": 1132, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1132, "pid": 2, "tid": 0, "ts": 7429564839476.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564839494.364, "dur": 3.968,
    "args": {
      "External id": 3087,      "device": 2, "stream": 0,      "correlation": 1133, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1133, "pid": 2, "tid": 0, "ts": 7429564839494.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839508.116, "dur": 3.487,
    "args": {
      "External id": 3097,      "device": 2, "stream": 0,      "correlation": 1134, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1134, "pid": 2, "tid": 0, "ts": 7429564839508.116,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839535.058, "dur": 3.567,
    "args": {
      "External id": 3102,      "device": 2, "stream": 0,      "correlation": 1135, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1135, "pid": 2, "tid": 0, "ts": 7429564839535.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564839556.748, "dur": 6.453,
    "args": {
      "External id": 3103,      "device": 2, "stream": 0,      "correlation": 1136, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1136, "pid": 2, "tid": 0, "ts": 7429564839556.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839614.079, "dur": 3.688,
    "args": {
      "External id": 3104,      "device": 2, "stream": 0,      "correlation": 1139, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1139, "pid": 2, "tid": 0, "ts": 7429564839614.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839627.711, "dur": 3.928,
    "args": {
      "External id": 3105,      "device": 2, "stream": 0,      "correlation": 1140, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1140, "pid": 2, "tid": 0, "ts": 7429564839627.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839637.573, "dur": 3.648,
    "args": {
      "External id": 3106,      "device": 2, "stream": 0,      "correlation": 1141, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1141, "pid": 2, "tid": 0, "ts": 7429564839637.573,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839656.176, "dur": 3.607,
    "args": {
      "External id": 3111,      "device": 2, "stream": 0,      "correlation": 1142, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1142, "pid": 2, "tid": 0, "ts": 7429564839656.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564839668.484, "dur": 6.093,
    "args": {
      "External id": 3112,      "device": 2, "stream": 0,      "correlation": 1143, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1143, "pid": 2, "tid": 0, "ts": 7429564839668.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839679.871, "dur": 3.687,
    "args": {
      "External id": 3113,      "device": 2, "stream": 0,      "correlation": 1144, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1144, "pid": 2, "tid": 0, "ts": 7429564839679.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839688.410, "dur": 3.968,
    "args": {
      "External id": 3114,      "device": 2, "stream": 0,      "correlation": 1145, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1145, "pid": 2, "tid": 0, "ts": 7429564839688.410,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839755.805, "dur": 3.728,
    "args": {
      "External id": 3131,      "device": 2, "stream": 0,      "correlation": 1146, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1146, "pid": 2, "tid": 0, "ts": 7429564839755.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839789.322, "dur": 4.169,
    "args": {
      "External id": 3149,      "device": 2, "stream": 0,      "correlation": 1147, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1147, "pid": 2, "tid": 0, "ts": 7429564839789.322,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564839842.204, "dur": 6.053,
    "args": {
      "External id": 3162,      "device": 2, "stream": 0,      "correlation": 1151, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1151, "pid": 2, "tid": 0, "ts": 7429564839842.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564839859.042, "dur": 4.249,
    "args": {
      "External id": 3165,      "device": 2, "stream": 0,      "correlation": 1152, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1152, "pid": 2, "tid": 0, "ts": 7429564839859.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839877.565, "dur": 3.687,
    "args": {
      "External id": 3172,      "device": 2, "stream": 0,      "correlation": 1153, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1153, "pid": 2, "tid": 0, "ts": 7429564839877.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564839895.246, "dur": 3.687,
    "args": {
      "External id": 3175,      "device": 2, "stream": 0,      "correlation": 1154, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1154, "pid": 2, "tid": 0, "ts": 7429564839895.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564839952.778, "dur": 5.652,
    "args": {
      "External id": 3187,      "device": 2, "stream": 0,      "correlation": 1158, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1158, "pid": 2, "tid": 0, "ts": 7429564839952.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564839972.223, "dur": 3.767,
    "args": {
      "External id": 3196,      "device": 2, "stream": 0,      "correlation": 1159, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1159, "pid": 2, "tid": 0, "ts": 7429564839972.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564840029.635, "dur": 68.476,
    "args": {
      "External id": 3206,      "device": 2, "stream": 0,      "correlation": 1163, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1163, "pid": 2, "tid": 0, "ts": 7429564840029.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564840098.112, "dur": 5.331,
    "args": {
      "External id": 3209,      "device": 2, "stream": 0,      "correlation": 1164, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1164, "pid": 2, "tid": 0, "ts": 7429564840098.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564840103.444, "dur": 4.930,
    "args": {
      "External id": 3211,      "device": 2, "stream": 0,      "correlation": 1165, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1165, "pid": 2, "tid": 0, "ts": 7429564840103.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564840108.375, "dur": 6.976,
    "args": {
      "External id": 3214,      "device": 2, "stream": 0,      "correlation": 1166, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1166, "pid": 2, "tid": 0, "ts": 7429564840108.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564840115.352, "dur": 4.529,
    "args": {
      "External id": 3215,      "device": 2, "stream": 0,      "correlation": 1167, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1167, "pid": 2, "tid": 0, "ts": 7429564840115.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564840119.882, "dur": 5.812,
    "args": {
      "External id": 3216,      "device": 2, "stream": 0,      "correlation": 1168, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1168, "pid": 2, "tid": 0, "ts": 7429564840119.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840125.695, "dur": 4.129,
    "args": {
      "External id": 3217,      "device": 2, "stream": 0,      "correlation": 1169, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1169, "pid": 2, "tid": 0, "ts": 7429564840125.695,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840135.358, "dur": 3.807,
    "args": {
      "External id": 3219,      "device": 2, "stream": 0,      "correlation": 1170, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1170, "pid": 2, "tid": 0, "ts": 7429564840135.358,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564840191.847, "dur": 81.627,
    "args": {
      "External id": 3227,      "device": 2, "stream": 0,      "correlation": 1174, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1174, "pid": 2, "tid": 0, "ts": 7429564840191.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564840273.475, "dur": 4.970,
    "args": {
      "External id": 3227,      "device": 2, "stream": 0,      "correlation": 1175, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1175, "pid": 2, "tid": 0, "ts": 7429564840273.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564840278.446, "dur": 4.049,
    "args": {
      "External id": 3230,      "device": 2, "stream": 0,      "correlation": 1178, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1178, "pid": 2, "tid": 0, "ts": 7429564840278.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564840314.329, "dur": 79.381,
    "args": {
      "External id": 3238,      "device": 2, "stream": 0,      "correlation": 1182, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1182, "pid": 2, "tid": 0, "ts": 7429564840314.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564840393.711, "dur": 4.690,
    "args": {
      "External id": 3238,      "device": 2, "stream": 0,      "correlation": 1183, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1183, "pid": 2, "tid": 0, "ts": 7429564840393.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564840398.402, "dur": 4.209,
    "args": {
      "External id": 3241,      "device": 2, "stream": 0,      "correlation": 1186, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1186, "pid": 2, "tid": 0, "ts": 7429564840398.402,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564840410.630, "dur": 88.042,
    "args": {
      "External id": 3249,      "device": 2, "stream": 0,      "correlation": 1190, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1190, "pid": 2, "tid": 0, "ts": 7429564840410.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564840498.673, "dur": 4.649,
    "args": {
      "External id": 3249,      "device": 2, "stream": 0,      "correlation": 1191, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1191, "pid": 2, "tid": 0, "ts": 7429564840498.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564840503.323, "dur": 4.730,
    "args": {
      "External id": 3252,      "device": 2, "stream": 0,      "correlation": 1192, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1192, "pid": 2, "tid": 0, "ts": 7429564840503.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564840508.054, "dur": 5.051,
    "args": {
      "External id": 3254,      "device": 2, "stream": 0,      "correlation": 1193, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1193, "pid": 2, "tid": 0, "ts": 7429564840508.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564840513.106, "dur": 7.175,
    "args": {
      "External id": 3257,      "device": 2, "stream": 0,      "correlation": 1194, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1194, "pid": 2, "tid": 0, "ts": 7429564840513.106,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564840520.282, "dur": 4.530,
    "args": {
      "External id": 3258,      "device": 2, "stream": 0,      "correlation": 1195, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1195, "pid": 2, "tid": 0, "ts": 7429564840520.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564840524.813, "dur": 5.531,
    "args": {
      "External id": 3259,      "device": 2, "stream": 0,      "correlation": 1196, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1196, "pid": 2, "tid": 0, "ts": 7429564840524.813,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840530.345, "dur": 4.810,
    "args": {
      "External id": 3260,      "device": 2, "stream": 0,      "correlation": 1197, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1197, "pid": 2, "tid": 0, "ts": 7429564840530.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840535.156, "dur": 4.490,
    "args": {
      "External id": 3262,      "device": 2, "stream": 0,      "correlation": 1198, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1198, "pid": 2, "tid": 0, "ts": 7429564840535.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564840579.057, "dur": 68.196,
    "args": {
      "External id": 3270,      "device": 2, "stream": 0,      "correlation": 1202, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1202, "pid": 2, "tid": 0, "ts": 7429564840579.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564840647.254, "dur": 15.395,
    "args": {
      "External id": 3283,      "device": 2, "stream": 0,      "correlation": 1206, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1206, "pid": 2, "tid": 0, "ts": 7429564840647.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564840662.650, "dur": 3.968,
    "args": {
      "External id": 3283,      "device": 2, "stream": 0,      "correlation": 1207, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1207, "pid": 2, "tid": 0, "ts": 7429564840662.650,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564840685.903, "dur": 17.118,
    "args": {
      "External id": 3296,      "device": 2, "stream": 0,      "correlation": 1211, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1211, "pid": 2, "tid": 0, "ts": 7429564840685.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564840703.022, "dur": 3.928,
    "args": {
      "External id": 3296,      "device": 2, "stream": 0,      "correlation": 1212, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1212, "pid": 2, "tid": 0, "ts": 7429564840703.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840717.656, "dur": 3.768,
    "args": {
      "External id": 3306,      "device": 2, "stream": 0,      "correlation": 1213, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1213, "pid": 2, "tid": 0, "ts": 7429564840717.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840748.166, "dur": 3.407,
    "args": {
      "External id": 3311,      "device": 2, "stream": 0,      "correlation": 1214, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1214, "pid": 2, "tid": 0, "ts": 7429564840748.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564840768.172, "dur": 6.855,
    "args": {
      "External id": 3312,      "device": 2, "stream": 0,      "correlation": 1215, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1215, "pid": 2, "tid": 0, "ts": 7429564840768.172,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840780.641, "dur": 3.567,
    "args": {
      "External id": 3313,      "device": 2, "stream": 0,      "correlation": 1216, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1216, "pid": 2, "tid": 0, "ts": 7429564840780.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840791.385, "dur": 3.568,
    "args": {
      "External id": 3314,      "device": 2, "stream": 0,      "correlation": 1217, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1217, "pid": 2, "tid": 0, "ts": 7429564840791.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840800.607, "dur": 3.567,
    "args": {
      "External id": 3315,      "device": 2, "stream": 0,      "correlation": 1218, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1218, "pid": 2, "tid": 0, "ts": 7429564840800.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840818.488, "dur": 3.807,
    "args": {
      "External id": 3320,      "device": 2, "stream": 0,      "correlation": 1219, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1219, "pid": 2, "tid": 0, "ts": 7429564840818.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564840832.400, "dur": 6.253,
    "args": {
      "External id": 3321,      "device": 2, "stream": 0,      "correlation": 1220, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1220, "pid": 2, "tid": 0, "ts": 7429564840832.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840842.984, "dur": 3.086,
    "args": {
      "External id": 3322,      "device": 2, "stream": 0,      "correlation": 1221, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1221, "pid": 2, "tid": 0, "ts": 7429564840842.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840849.920, "dur": 3.687,
    "args": {
      "External id": 3323,      "device": 2, "stream": 0,      "correlation": 1222, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1222, "pid": 2, "tid": 0, "ts": 7429564840849.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840912.504, "dur": 3.808,
    "args": {
      "External id": 3340,      "device": 2, "stream": 0,      "correlation": 1223, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1223, "pid": 2, "tid": 0, "ts": 7429564840912.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564840946.141, "dur": 3.607,
    "args": {
      "External id": 3358,      "device": 2, "stream": 0,      "correlation": 1224, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1224, "pid": 2, "tid": 0, "ts": 7429564840946.141,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564840993.049, "dur": 6.053,
    "args": {
      "External id": 3371,      "device": 2, "stream": 0,      "correlation": 1228, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1228, "pid": 2, "tid": 0, "ts": 7429564840993.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841010.008, "dur": 3.367,
    "args": {
      "External id": 3374,      "device": 2, "stream": 0,      "correlation": 1229, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1229, "pid": 2, "tid": 0, "ts": 7429564841010.008,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841027.809, "dur": 3.086,
    "args": {
      "External id": 3381,      "device": 2, "stream": 0,      "correlation": 1230, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1230, "pid": 2, "tid": 0, "ts": 7429564841027.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564841045.169, "dur": 2.845,
    "args": {
      "External id": 3384,      "device": 2, "stream": 0,      "correlation": 1231, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1231, "pid": 2, "tid": 0, "ts": 7429564841045.169,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564841102.661, "dur": 4.770,
    "args": {
      "External id": 3396,      "device": 2, "stream": 0,      "correlation": 1235, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1235, "pid": 2, "tid": 0, "ts": 7429564841102.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841123.830, "dur": 2.965,
    "args": {
      "External id": 3405,      "device": 2, "stream": 0,      "correlation": 1236, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1236, "pid": 2, "tid": 0, "ts": 7429564841123.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564841177.994, "dur": 66.833,
    "args": {
      "External id": 3415,      "device": 2, "stream": 0,      "correlation": 1240, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1240, "pid": 2, "tid": 0, "ts": 7429564841177.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841244.828, "dur": 4.770,
    "args": {
      "External id": 3418,      "device": 2, "stream": 0,      "correlation": 1241, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1241, "pid": 2, "tid": 0, "ts": 7429564841244.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841249.599, "dur": 5.010,
    "args": {
      "External id": 3420,      "device": 2, "stream": 0,      "correlation": 1242, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1242, "pid": 2, "tid": 0, "ts": 7429564841249.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564841254.610, "dur": 7.577,
    "args": {
      "External id": 3423,      "device": 2, "stream": 0,      "correlation": 1243, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1243, "pid": 2, "tid": 0, "ts": 7429564841254.610,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841262.188, "dur": 5.812,
    "args": {
      "External id": 3424,      "device": 2, "stream": 0,      "correlation": 1244, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1244, "pid": 2, "tid": 0, "ts": 7429564841262.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841268.001, "dur": 4.730,
    "args": {
      "External id": 3425,      "device": 2, "stream": 0,      "correlation": 1245, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1245, "pid": 2, "tid": 0, "ts": 7429564841268.001,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841272.732, "dur": 6.414,
    "args": {
      "External id": 3426,      "device": 2, "stream": 0,      "correlation": 1246, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1246, "pid": 2, "tid": 0, "ts": 7429564841272.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841279.147, "dur": 4.128,
    "args": {
      "External id": 3428,      "device": 2, "stream": 0,      "correlation": 1247, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1247, "pid": 2, "tid": 0, "ts": 7429564841279.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564841395.454, "dur": 76.254,
    "args": {
      "External id": 3436,      "device": 2, "stream": 0,      "correlation": 1253, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1253, "pid": 2, "tid": 0, "ts": 7429564841395.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564841471.709, "dur": 5.011,
    "args": {
      "External id": 3436,      "device": 2, "stream": 0,      "correlation": 1254, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1254, "pid": 2, "tid": 0, "ts": 7429564841471.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841476.721, "dur": 4.048,
    "args": {
      "External id": 3439,      "device": 2, "stream": 0,      "correlation": 1255, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1255, "pid": 2, "tid": 0, "ts": 7429564841476.721,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564841504.705, "dur": 76.615,
    "args": {
      "External id": 3447,      "device": 2, "stream": 0,      "correlation": 1261, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1261, "pid": 2, "tid": 0, "ts": 7429564841504.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564841581.321, "dur": 5.091,
    "args": {
      "External id": 3447,      "device": 2, "stream": 0,      "correlation": 1262, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1262, "pid": 2, "tid": 0, "ts": 7429564841581.321,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841586.413, "dur": 4.970,
    "args": {
      "External id": 3450,      "device": 2, "stream": 0,      "correlation": 1263, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1263, "pid": 2, "tid": 0, "ts": 7429564841586.413,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564841591.384, "dur": 87.280,
    "args": {
      "External id": 3458,      "device": 2, "stream": 0,      "correlation": 1267, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1267, "pid": 2, "tid": 0, "ts": 7429564841591.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564841678.665, "dur": 4.690,
    "args": {
      "External id": 3458,      "device": 2, "stream": 0,      "correlation": 1268, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1268, "pid": 2, "tid": 0, "ts": 7429564841678.665,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841683.356, "dur": 4.649,
    "args": {
      "External id": 3461,      "device": 2, "stream": 0,      "correlation": 1269, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1269, "pid": 2, "tid": 0, "ts": 7429564841683.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841688.006, "dur": 4.570,
    "args": {
      "External id": 3463,      "device": 2, "stream": 0,      "correlation": 1270, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1270, "pid": 2, "tid": 0, "ts": 7429564841688.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564841692.577, "dur": 7.095,
    "args": {
      "External id": 3466,      "device": 2, "stream": 0,      "correlation": 1271, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1271, "pid": 2, "tid": 0, "ts": 7429564841692.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841699.673, "dur": 4.530,
    "args": {
      "External id": 3467,      "device": 2, "stream": 0,      "correlation": 1272, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1272, "pid": 2, "tid": 0, "ts": 7429564841699.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564841704.204, "dur": 4.529,
    "args": {
      "External id": 3468,      "device": 2, "stream": 0,      "correlation": 1273, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1273, "pid": 2, "tid": 0, "ts": 7429564841704.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841708.734, "dur": 4.690,
    "args": {
      "External id": 3469,      "device": 2, "stream": 0,      "correlation": 1274, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1274, "pid": 2, "tid": 0, "ts": 7429564841708.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841713.425, "dur": 4.128,
    "args": {
      "External id": 3471,      "device": 2, "stream": 0,      "correlation": 1275, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1275, "pid": 2, "tid": 0, "ts": 7429564841713.425,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564841727.698, "dur": 67.273,
    "args": {
      "External id": 3479,      "device": 2, "stream": 0,      "correlation": 1279, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1279, "pid": 2, "tid": 0, "ts": 7429564841727.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564841794.972, "dur": 15.635,
    "args": {
      "External id": 3492,      "device": 2, "stream": 0,      "correlation": 1283, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1283, "pid": 2, "tid": 0, "ts": 7429564841794.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564841810.608, "dur": 4.169,
    "args": {
      "External id": 3492,      "device": 2, "stream": 0,      "correlation": 1284, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1284, "pid": 2, "tid": 0, "ts": 7429564841810.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564841827.527, "dur": 16.277,
    "args": {
      "External id": 3505,      "device": 2, "stream": 0,      "correlation": 1288, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1288, "pid": 2, "tid": 0, "ts": 7429564841827.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564841843.805, "dur": 3.968,
    "args": {
      "External id": 3505,      "device": 2, "stream": 0,      "correlation": 1289, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1289, "pid": 2, "tid": 0, "ts": 7429564841843.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841894.481, "dur": 3.808,
    "args": {
      "External id": 3515,      "device": 2, "stream": 0,      "correlation": 1292, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1292, "pid": 2, "tid": 0, "ts": 7429564841894.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841924.510, "dur": 3.567,
    "args": {
      "External id": 3520,      "device": 2, "stream": 0,      "correlation": 1293, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1293, "pid": 2, "tid": 0, "ts": 7429564841924.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564841947.042, "dur": 6.614,
    "args": {
      "External id": 3521,      "device": 2, "stream": 0,      "correlation": 1294, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1294, "pid": 2, "tid": 0, "ts": 7429564841947.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841959.430, "dur": 3.688,
    "args": {
      "External id": 3522,      "device": 2, "stream": 0,      "correlation": 1295, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1295, "pid": 2, "tid": 0, "ts": 7429564841959.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841969.052, "dur": 3.888,
    "args": {
      "External id": 3523,      "device": 2, "stream": 0,      "correlation": 1296, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1296, "pid": 2, "tid": 0, "ts": 7429564841969.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841978.634, "dur": 3.528,
    "args": {
      "External id": 3524,      "device": 2, "stream": 0,      "correlation": 1297, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1297, "pid": 2, "tid": 0, "ts": 7429564841978.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564841996.836, "dur": 3.126,
    "args": {
      "External id": 3529,      "device": 2, "stream": 0,      "correlation": 1298, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1298, "pid": 2, "tid": 0, "ts": 7429564841996.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564842010.468, "dur": 6.253,
    "args": {
      "External id": 3530,      "device": 2, "stream": 0,      "correlation": 1299, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1299, "pid": 2, "tid": 0, "ts": 7429564842010.468,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564842021.373, "dur": 2.364,
    "args": {
      "External id": 3531,      "device": 2, "stream": 0,      "correlation": 1300, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1300, "pid": 2, "tid": 0, "ts": 7429564842021.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564842027.787, "dur": 3.928,
    "args": {
      "External id": 3532,      "device": 2, "stream": 0,      "correlation": 1301, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1301, "pid": 2, "tid": 0, "ts": 7429564842027.787,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564842116.030, "dur": 4.008,
    "args": {
      "External id": 3549,      "device": 2, "stream": 0,      "correlation": 1302, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1302, "pid": 2, "tid": 0, "ts": 7429564842116.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564842156.644, "dur": 3.888,
    "args": {
      "External id": 3567,      "device": 2, "stream": 0,      "correlation": 1303, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1303, "pid": 2, "tid": 0, "ts": 7429564842156.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564842215.619, "dur": 5.973,
    "args": {
      "External id": 3580,      "device": 2, "stream": 0,      "correlation": 1307, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1307, "pid": 2, "tid": 0, "ts": 7429564842215.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842233.781, "dur": 4.048,
    "args": {
      "External id": 3583,      "device": 2, "stream": 0,      "correlation": 1308, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1308, "pid": 2, "tid": 0, "ts": 7429564842233.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564842291.193, "dur": 3.487,
    "args": {
      "External id": 3590,      "device": 2, "stream": 0,      "correlation": 1309, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1309, "pid": 2, "tid": 0, "ts": 7429564842291.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564842309.916, "dur": 3.567,
    "args": {
      "External id": 3593,      "device": 2, "stream": 0,      "correlation": 1310, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1310, "pid": 2, "tid": 0, "ts": 7429564842309.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564842366.005, "dur": 5.371,
    "args": {
      "External id": 3605,      "device": 2, "stream": 0,      "correlation": 1314, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1314, "pid": 2, "tid": 0, "ts": 7429564842366.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564842384.888, "dur": 4.089,
    "args": {
      "External id": 3614,      "device": 2, "stream": 0,      "correlation": 1315, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1315, "pid": 2, "tid": 0, "ts": 7429564842384.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564842442.420, "dur": 67.916,
    "args": {
      "External id": 3624,      "device": 2, "stream": 0,      "correlation": 1319, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1319, "pid": 2, "tid": 0, "ts": 7429564842442.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842510.337, "dur": 4.770,
    "args": {
      "External id": 3627,      "device": 2, "stream": 0,      "correlation": 1320, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1320, "pid": 2, "tid": 0, "ts": 7429564842510.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842515.108, "dur": 4.850,
    "args": {
      "External id": 3629,      "device": 2, "stream": 0,      "correlation": 1321, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1321, "pid": 2, "tid": 0, "ts": 7429564842515.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564842519.959, "dur": 6.975,
    "args": {
      "External id": 3632,      "device": 2, "stream": 0,      "correlation": 1322, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1322, "pid": 2, "tid": 0, "ts": 7429564842519.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842526.935, "dur": 4.529,
    "args": {
      "External id": 3633,      "device": 2, "stream": 0,      "correlation": 1323, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1323, "pid": 2, "tid": 0, "ts": 7429564842526.935,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842531.465, "dur": 4.530,
    "args": {
      "External id": 3634,      "device": 2, "stream": 0,      "correlation": 1324, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1324, "pid": 2, "tid": 0, "ts": 7429564842531.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564842535.996, "dur": 4.248,
    "args": {
      "External id": 3635,      "device": 2, "stream": 0,      "correlation": 1325, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1325, "pid": 2, "tid": 0, "ts": 7429564842535.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564842544.255, "dur": 3.647,
    "args": {
      "External id": 3637,      "device": 2, "stream": 0,      "correlation": 1326, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1326, "pid": 2, "tid": 0, "ts": 7429564842544.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564842598.780, "dur": 91.369,
    "args": {
      "External id": 3645,      "device": 2, "stream": 0,      "correlation": 1330, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1330, "pid": 2, "tid": 0, "ts": 7429564842598.780,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564842690.150, "dur": 4.810,
    "args": {
      "External id": 3645,      "device": 2, "stream": 0,      "correlation": 1331, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1331, "pid": 2, "tid": 0, "ts": 7429564842690.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842694.961, "dur": 4.128,
    "args": {
      "External id": 3648,      "device": 2, "stream": 0,      "correlation": 1334, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1334, "pid": 2, "tid": 0, "ts": 7429564842694.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564842731.445, "dur": 84.433,
    "args": {
      "External id": 3656,      "device": 2, "stream": 0,      "correlation": 1338, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1338, "pid": 2, "tid": 0, "ts": 7429564842731.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564842815.879, "dur": 4.770,
    "args": {
      "External id": 3656,      "device": 2, "stream": 0,      "correlation": 1339, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1339, "pid": 2, "tid": 0, "ts": 7429564842815.879,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842820.650, "dur": 4.730,
    "args": {
      "External id": 3659,      "device": 2, "stream": 0,      "correlation": 1342, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1342, "pid": 2, "tid": 0, "ts": 7429564842820.650,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564842825.381, "dur": 86.397,
    "args": {
      "External id": 3667,      "device": 2, "stream": 0,      "correlation": 1346, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1346, "pid": 2, "tid": 0, "ts": 7429564842825.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564842911.779, "dur": 4.570,
    "args": {
      "External id": 3667,      "device": 2, "stream": 0,      "correlation": 1347, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1347, "pid": 2, "tid": 0, "ts": 7429564842911.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842916.350, "dur": 4.569,
    "args": {
      "External id": 3670,      "device": 2, "stream": 0,      "correlation": 1348, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1348, "pid": 2, "tid": 0, "ts": 7429564842916.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842920.920, "dur": 5.813,
    "args": {
      "External id": 3672,      "device": 2, "stream": 0,      "correlation": 1349, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1349, "pid": 2, "tid": 0, "ts": 7429564842920.920,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564842926.734, "dur": 7.616,
    "args": {
      "External id": 3675,      "device": 2, "stream": 0,      "correlation": 1350, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1350, "pid": 2, "tid": 0, "ts": 7429564842926.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842934.351, "dur": 5.772,
    "args": {
      "External id": 3676,      "device": 2, "stream": 0,      "correlation": 1351, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1351, "pid": 2, "tid": 0, "ts": 7429564842934.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564842940.124, "dur": 4.730,
    "args": {
      "External id": 3677,      "device": 2, "stream": 0,      "correlation": 1352, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1352, "pid": 2, "tid": 0, "ts": 7429564842940.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564842944.855, "dur": 4.730,
    "args": {
      "External id": 3678,      "device": 2, "stream": 0,      "correlation": 1353, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1353, "pid": 2, "tid": 0, "ts": 7429564842944.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564842949.586, "dur": 4.089,
    "args": {
      "External id": 3680,      "device": 2, "stream": 0,      "correlation": 1354, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1354, "pid": 2, "tid": 0, "ts": 7429564842949.586,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564842987.634, "dur": 68.436,
    "args": {
      "External id": 3688,      "device": 2, "stream": 0,      "correlation": 1358, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1358, "pid": 2, "tid": 0, "ts": 7429564842987.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564843056.071, "dur": 15.475,
    "args": {
      "External id": 3701,      "device": 2, "stream": 0,      "correlation": 1362, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1362, "pid": 2, "tid": 0, "ts": 7429564843056.071,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564843071.547, "dur": 3.928,
    "args": {
      "External id": 3701,      "device": 2, "stream": 0,      "correlation": 1363, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1363, "pid": 2, "tid": 0, "ts": 7429564843071.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564843090.149, "dur": 15.916,
    "args": {
      "External id": 3714,      "device": 2, "stream": 0,      "correlation": 1367, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1367, "pid": 2, "tid": 0, "ts": 7429564843090.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564843106.066, "dur": 4.008,
    "args": {
      "External id": 3714,      "device": 2, "stream": 0,      "correlation": 1368, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1368, "pid": 2, "tid": 0, "ts": 7429564843106.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843123.747, "dur": 2.524,
    "args": {
      "External id": 3724,      "device": 2, "stream": 0,      "correlation": 1369, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1369, "pid": 2, "tid": 0, "ts": 7429564843123.747,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843151.009, "dur": 3.167,
    "args": {
      "External id": 3729,      "device": 2, "stream": 0,      "correlation": 1370, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1370, "pid": 2, "tid": 0, "ts": 7429564843151.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564843172.900, "dur": 6.213,
    "args": {
      "External id": 3730,      "device": 2, "stream": 0,      "correlation": 1371, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1371, "pid": 2, "tid": 0, "ts": 7429564843172.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843184.406, "dur": 2.725,
    "args": {
      "External id": 3731,      "device": 2, "stream": 0,      "correlation": 1372, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1372, "pid": 2, "tid": 0, "ts": 7429564843184.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843192.946, "dur": 3.046,
    "args": {
      "External id": 3732,      "device": 2, "stream": 0,      "correlation": 1373, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1373, "pid": 2, "tid": 0, "ts": 7429564843192.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843202.929, "dur": 2.765,
    "args": {
      "External id": 3733,      "device": 2, "stream": 0,      "correlation": 1374, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1374, "pid": 2, "tid": 0, "ts": 7429564843202.929,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843220.810, "dur": 2.765,
    "args": {
      "External id": 3738,      "device": 2, "stream": 0,      "correlation": 1375, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1375, "pid": 2, "tid": 0, "ts": 7429564843220.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564843234.120, "dur": 5.171,
    "args": {
      "External id": 3739,      "device": 2, "stream": 0,      "correlation": 1376, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1376, "pid": 2, "tid": 0, "ts": 7429564843234.120,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843245.146, "dur": 3.046,
    "args": {
      "External id": 3740,      "device": 2, "stream": 0,      "correlation": 1377, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1377, "pid": 2, "tid": 0, "ts": 7429564843245.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843254.166, "dur": 3.447,
    "args": {
      "External id": 3741,      "device": 2, "stream": 0,      "correlation": 1378, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1378, "pid": 2, "tid": 0, "ts": 7429564843254.166,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843317.993, "dur": 3.808,
    "args": {
      "External id": 3758,      "device": 2, "stream": 0,      "correlation": 1379, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1379, "pid": 2, "tid": 0, "ts": 7429564843317.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843354.116, "dur": 2.926,
    "args": {
      "External id": 3776,      "device": 2, "stream": 0,      "correlation": 1380, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1380, "pid": 2, "tid": 0, "ts": 7429564843354.116,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564843406.797, "dur": 5.412,
    "args": {
      "External id": 3789,      "device": 2, "stream": 0,      "correlation": 1384, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1384, "pid": 2, "tid": 0, "ts": 7429564843406.797,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564843421.952, "dur": 3.246,
    "args": {
      "External id": 3792,      "device": 2, "stream": 0,      "correlation": 1385, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1385, "pid": 2, "tid": 0, "ts": 7429564843421.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843440.675, "dur": 2.926,
    "args": {
      "External id": 3799,      "device": 2, "stream": 0,      "correlation": 1386, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1386, "pid": 2, "tid": 0, "ts": 7429564843440.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564843458.436, "dur": 3.166,
    "args": {
      "External id": 3802,      "device": 2, "stream": 0,      "correlation": 1387, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1387, "pid": 2, "tid": 0, "ts": 7429564843458.436,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564843512.721, "dur": 4.810,
    "args": {
      "External id": 3814,      "device": 2, "stream": 0,      "correlation": 1391, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1391, "pid": 2, "tid": 0, "ts": 7429564843512.721,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843531.564, "dur": 3.928,
    "args": {
      "External id": 3823,      "device": 2, "stream": 0,      "correlation": 1392, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1392, "pid": 2, "tid": 0, "ts": 7429564843531.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564843586.450, "dur": 68.837,
    "args": {
      "External id": 3833,      "device": 2, "stream": 0,      "correlation": 1396, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1396, "pid": 2, "tid": 0, "ts": 7429564843586.450,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564843659.458, "dur": 3.166,
    "args": {
      "External id": 3836,      "device": 2, "stream": 0,      "correlation": 1399, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1399, "pid": 2, "tid": 0, "ts": 7429564843659.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564843687.643, "dur": 3.046,
    "args": {
      "External id": 3838,      "device": 2, "stream": 0,      "correlation": 1400, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1400, "pid": 2, "tid": 0, "ts": 7429564843687.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564843705.604, "dur": 8.899,
    "args": {
      "External id": 3841,      "device": 2, "stream": 0,      "correlation": 1401, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1401, "pid": 2, "tid": 0, "ts": 7429564843705.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564843718.233, "dur": 4.129,
    "args": {
      "External id": 3842,      "device": 2, "stream": 0,      "correlation": 1402, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1402, "pid": 2, "tid": 0, "ts": 7429564843718.233,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564843729.258, "dur": 4.410,
    "args": {
      "External id": 3843,      "device": 2, "stream": 0,      "correlation": 1403, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1403, "pid": 2, "tid": 0, "ts": 7429564843729.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843741.847, "dur": 3.768,
    "args": {
      "External id": 3844,      "device": 2, "stream": 0,      "correlation": 1404, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1404, "pid": 2, "tid": 0, "ts": 7429564843741.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564843754.957, "dur": 3.528,
    "args": {
      "External id": 3846,      "device": 2, "stream": 0,      "correlation": 1405, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1405, "pid": 2, "tid": 0, "ts": 7429564843754.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564843816.218, "dur": 83.431,
    "args": {
      "External id": 3854,      "device": 2, "stream": 0,      "correlation": 1409, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1409, "pid": 2, "tid": 0, "ts": 7429564843816.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564843899.650, "dur": 4.810,
    "args": {
      "External id": 3854,      "device": 2, "stream": 0,      "correlation": 1410, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1410, "pid": 2, "tid": 0, "ts": 7429564843899.650,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564843904.461, "dur": 4.169,
    "args": {
      "External id": 3857,      "device": 2, "stream": 0,      "correlation": 1411, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1411, "pid": 2, "tid": 0, "ts": 7429564843904.461,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564843929.599, "dur": 88.763,
    "args": {
      "External id": 3865,      "device": 2, "stream": 0,      "correlation": 1417, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1417, "pid": 2, "tid": 0, "ts": 7429564843929.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564844018.363, "dur": 5.051,
    "args": {
      "External id": 3865,      "device": 2, "stream": 0,      "correlation": 1418, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1418, "pid": 2, "tid": 0, "ts": 7429564844018.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564844023.415, "dur": 4.970,
    "args": {
      "External id": 3868,      "device": 2, "stream": 0,      "correlation": 1419, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1419, "pid": 2, "tid": 0, "ts": 7429564844023.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564844028.386, "dur": 90.888,
    "args": {
      "External id": 3876,      "device": 2, "stream": 0,      "correlation": 1423, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1423, "pid": 2, "tid": 0, "ts": 7429564844028.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564844119.275, "dur": 4.690,
    "args": {
      "External id": 3876,      "device": 2, "stream": 0,      "correlation": 1424, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1424, "pid": 2, "tid": 0, "ts": 7429564844119.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564844123.966, "dur": 4.810,
    "args": {
      "External id": 3879,      "device": 2, "stream": 0,      "correlation": 1425, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1425, "pid": 2, "tid": 0, "ts": 7429564844123.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564844128.777, "dur": 4.770,
    "args": {
      "External id": 3881,      "device": 2, "stream": 0,      "correlation": 1426, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1426, "pid": 2, "tid": 0, "ts": 7429564844128.777,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564844133.548, "dur": 7.095,
    "args": {
      "External id": 3884,      "device": 2, "stream": 0,      "correlation": 1427, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1427, "pid": 2, "tid": 0, "ts": 7429564844133.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564844140.644, "dur": 4.569,
    "args": {
      "External id": 3885,      "device": 2, "stream": 0,      "correlation": 1428, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1428, "pid": 2, "tid": 0, "ts": 7429564844140.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564844145.214, "dur": 4.730,
    "args": {
      "External id": 3886,      "device": 2, "stream": 0,      "correlation": 1429, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1429, "pid": 2, "tid": 0, "ts": 7429564844145.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844149.945, "dur": 4.690,
    "args": {
      "External id": 3887,      "device": 2, "stream": 0,      "correlation": 1430, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1430, "pid": 2, "tid": 0, "ts": 7429564844149.945,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844154.636, "dur": 4.048,
    "args": {
      "External id": 3889,      "device": 2, "stream": 0,      "correlation": 1433, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1433, "pid": 2, "tid": 0, "ts": 7429564844154.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564844217.340, "dur": 67.996,
    "args": {
      "External id": 3897,      "device": 2, "stream": 0,      "correlation": 1437, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1437, "pid": 2, "tid": 0, "ts": 7429564844217.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564844285.337, "dur": 15.675,
    "args": {
      "External id": 3910,      "device": 2, "stream": 0,      "correlation": 1441, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1441, "pid": 2, "tid": 0, "ts": 7429564844285.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564844301.013, "dur": 4.008,
    "args": {
      "External id": 3910,      "device": 2, "stream": 0,      "correlation": 1442, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1442, "pid": 2, "tid": 0, "ts": 7429564844301.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564844327.113, "dur": 17.118,
    "args": {
      "External id": 3923,      "device": 2, "stream": 0,      "correlation": 1446, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1446, "pid": 2, "tid": 0, "ts": 7429564844327.113,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564844344.232, "dur": 3.888,
    "args": {
      "External id": 3923,      "device": 2, "stream": 0,      "correlation": 1447, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1447, "pid": 2, "tid": 0, "ts": 7429564844344.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844358.906, "dur": 4.008,
    "args": {
      "External id": 3933,      "device": 2, "stream": 0,      "correlation": 1448, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1448, "pid": 2, "tid": 0, "ts": 7429564844358.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844389.055, "dur": 3.287,
    "args": {
      "External id": 3938,      "device": 2, "stream": 0,      "correlation": 1449, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1449, "pid": 2, "tid": 0, "ts": 7429564844389.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564844409.382, "dur": 6.654,
    "args": {
      "External id": 3939,      "device": 2, "stream": 0,      "correlation": 1450, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1450, "pid": 2, "tid": 0, "ts": 7429564844409.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844422.813, "dur": 3.487,
    "args": {
      "External id": 3940,      "device": 2, "stream": 0,      "correlation": 1451, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1451, "pid": 2, "tid": 0, "ts": 7429564844422.813,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844435.442, "dur": 3.968,
    "args": {
      "External id": 3941,      "device": 2, "stream": 0,      "correlation": 1452, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1452, "pid": 2, "tid": 0, "ts": 7429564844435.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844445.104, "dur": 3.326,
    "args": {
      "External id": 3942,      "device": 2, "stream": 0,      "correlation": 1453, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1453, "pid": 2, "tid": 0, "ts": 7429564844445.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844463.225, "dur": 2.525,
    "args": {
      "External id": 3947,      "device": 2, "stream": 0,      "correlation": 1454, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1454, "pid": 2, "tid": 0, "ts": 7429564844463.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564844477.298, "dur": 5.411,
    "args": {
      "External id": 3948,      "device": 2, "stream": 0,      "correlation": 1455, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1455, "pid": 2, "tid": 0, "ts": 7429564844477.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844487.160, "dur": 2.646,
    "args": {
      "External id": 3949,      "device": 2, "stream": 0,      "correlation": 1456, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1456, "pid": 2, "tid": 0, "ts": 7429564844487.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844494.618, "dur": 3.687,
    "args": {
      "External id": 3950,      "device": 2, "stream": 0,      "correlation": 1457, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1457, "pid": 2, "tid": 0, "ts": 7429564844494.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844562.614, "dur": 3.046,
    "args": {
      "External id": 3967,      "device": 2, "stream": 0,      "correlation": 1458, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1458, "pid": 2, "tid": 0, "ts": 7429564844562.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844597.815, "dur": 3.367,
    "args": {
      "External id": 3985,      "device": 2, "stream": 0,      "correlation": 1459, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1459, "pid": 2, "tid": 0, "ts": 7429564844597.815,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564844646.527, "dur": 5.251,
    "args": {
      "External id": 3998,      "device": 2, "stream": 0,      "correlation": 1463, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1463, "pid": 2, "tid": 0, "ts": 7429564844646.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564844663.486, "dur": 3.447,
    "args": {
      "External id": 4001,      "device": 2, "stream": 0,      "correlation": 1464, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1464, "pid": 2, "tid": 0, "ts": 7429564844663.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844680.966, "dur": 3.006,
    "args": {
      "External id": 4008,      "device": 2, "stream": 0,      "correlation": 1465, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1465, "pid": 2, "tid": 0, "ts": 7429564844680.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564844697.965, "dur": 3.006,
    "args": {
      "External id": 4011,      "device": 2, "stream": 0,      "correlation": 1466, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1466, "pid": 2, "tid": 0, "ts": 7429564844697.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564844750.887, "dur": 5.171,
    "args": {
      "External id": 4023,      "device": 2, "stream": 0,      "correlation": 1470, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1470, "pid": 2, "tid": 0, "ts": 7429564844750.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844771.253, "dur": 3.046,
    "args": {
      "External id": 4032,      "device": 2, "stream": 0,      "correlation": 1471, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1471, "pid": 2, "tid": 0, "ts": 7429564844771.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564844821.248, "dur": 67.194,
    "args": {
      "External id": 4042,      "device": 2, "stream": 0,      "correlation": 1475, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1475, "pid": 2, "tid": 0, "ts": 7429564844821.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564844888.443, "dur": 5.050,
    "args": {
      "External id": 4045,      "device": 2, "stream": 0,      "correlation": 1476, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1476, "pid": 2, "tid": 0, "ts": 7429564844888.443,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564844893.494, "dur": 4.891,
    "args": {
      "External id": 4047,      "device": 2, "stream": 0,      "correlation": 1477, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1477, "pid": 2, "tid": 0, "ts": 7429564844893.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564844898.386, "dur": 7.255,
    "args": {
      "External id": 4050,      "device": 2, "stream": 0,      "correlation": 1478, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1478, "pid": 2, "tid": 0, "ts": 7429564844898.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564844905.642, "dur": 4.490,
    "args": {
      "External id": 4051,      "device": 2, "stream": 0,      "correlation": 1479, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1479, "pid": 2, "tid": 0, "ts": 7429564844905.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564844910.133, "dur": 5.131,
    "args": {
      "External id": 4052,      "device": 2, "stream": 0,      "correlation": 1480, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1480, "pid": 2, "tid": 0, "ts": 7429564844910.133,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844915.265, "dur": 4.288,
    "args": {
      "External id": 4053,      "device": 2, "stream": 0,      "correlation": 1481, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1481, "pid": 2, "tid": 0, "ts": 7429564844915.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564844926.049, "dur": 3.207,
    "args": {
      "External id": 4055,      "device": 2, "stream": 0,      "correlation": 1482, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1482, "pid": 2, "tid": 0, "ts": 7429564844926.049,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564845036.944, "dur": 83.110,
    "args": {
      "External id": 4063,      "device": 2, "stream": 0,      "correlation": 1488, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1488, "pid": 2, "tid": 0, "ts": 7429564845036.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564845120.055, "dur": 4.931,
    "args": {
      "External id": 4063,      "device": 2, "stream": 0,      "correlation": 1489, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1489, "pid": 2, "tid": 0, "ts": 7429564845120.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564845124.987, "dur": 4.048,
    "args": {
      "External id": 4066,      "device": 2, "stream": 0,      "correlation": 1490, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1490, "pid": 2, "tid": 0, "ts": 7429564845124.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564845144.712, "dur": 79.021,
    "args": {
      "External id": 4074,      "device": 2, "stream": 0,      "correlation": 1496, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1496, "pid": 2, "tid": 0, "ts": 7429564845144.712,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564845223.734, "dur": 4.810,
    "args": {
      "External id": 4074,      "device": 2, "stream": 0,      "correlation": 1497, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1497, "pid": 2, "tid": 0, "ts": 7429564845223.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564845228.545, "dur": 4.930,
    "args": {
      "External id": 4077,      "device": 2, "stream": 0,      "correlation": 1498, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1498, "pid": 2, "tid": 0, "ts": 7429564845228.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564845233.476, "dur": 87.199,
    "args": {
      "External id": 4085,      "device": 2, "stream": 0,      "correlation": 1502, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1502, "pid": 2, "tid": 0, "ts": 7429564845233.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564845320.676, "dur": 4.690,
    "args": {
      "External id": 4085,      "device": 2, "stream": 0,      "correlation": 1503, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1503, "pid": 2, "tid": 0, "ts": 7429564845320.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564845325.367, "dur": 4.730,
    "args": {
      "External id": 4088,      "device": 2, "stream": 0,      "correlation": 1504, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1504, "pid": 2, "tid": 0, "ts": 7429564845325.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564845330.098, "dur": 4.529,
    "args": {
      "External id": 4090,      "device": 2, "stream": 0,      "correlation": 1505, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1505, "pid": 2, "tid": 0, "ts": 7429564845330.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564845334.628, "dur": 7.216,
    "args": {
      "External id": 4093,      "device": 2, "stream": 0,      "correlation": 1506, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1506, "pid": 2, "tid": 0, "ts": 7429564845334.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564845341.845, "dur": 4.529,
    "args": {
      "External id": 4094,      "device": 2, "stream": 0,      "correlation": 1507, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1507, "pid": 2, "tid": 0, "ts": 7429564845341.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564845346.375, "dur": 4.610,
    "args": {
      "External id": 4095,      "device": 2, "stream": 0,      "correlation": 1508, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1508, "pid": 2, "tid": 0, "ts": 7429564845346.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845350.986, "dur": 5.572,
    "args": {
      "External id": 4096,      "device": 2, "stream": 0,      "correlation": 1509, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1509, "pid": 2, "tid": 0, "ts": 7429564845350.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845356.559, "dur": 4.128,
    "args": {
      "External id": 4098,      "device": 2, "stream": 0,      "correlation": 1510, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1510, "pid": 2, "tid": 0, "ts": 7429564845356.559,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564845406.233, "dur": 68.156,
    "args": {
      "External id": 4106,      "device": 2, "stream": 0,      "correlation": 1514, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1514, "pid": 2, "tid": 0, "ts": 7429564845406.233,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564845474.390, "dur": 15.635,
    "args": {
      "External id": 4119,      "device": 2, "stream": 0,      "correlation": 1518, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1518, "pid": 2, "tid": 0, "ts": 7429564845474.390,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564845490.026, "dur": 3.968,
    "args": {
      "External id": 4119,      "device": 2, "stream": 0,      "correlation": 1519, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1519, "pid": 2, "tid": 0, "ts": 7429564845490.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564845529.356, "dur": 17.038,
    "args": {
      "External id": 4132,      "device": 2, "stream": 0,      "correlation": 1523, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1523, "pid": 2, "tid": 0, "ts": 7429564845529.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564845546.395, "dur": 4.049,
    "args": {
      "External id": 4132,      "device": 2, "stream": 0,      "correlation": 1524, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1524, "pid": 2, "tid": 0, "ts": 7429564845546.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845560.989, "dur": 3.567,
    "args": {
      "External id": 4142,      "device": 2, "stream": 0,      "correlation": 1525, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1525, "pid": 2, "tid": 0, "ts": 7429564845560.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845588.011, "dur": 3.727,
    "args": {
      "External id": 4147,      "device": 2, "stream": 0,      "correlation": 1526, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1526, "pid": 2, "tid": 0, "ts": 7429564845588.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564845651.918, "dur": 6.814,
    "args": {
      "External id": 4148,      "device": 2, "stream": 0,      "correlation": 1529, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1529, "pid": 2, "tid": 0, "ts": 7429564845651.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845664.467, "dur": 3.567,
    "args": {
      "External id": 4149,      "device": 2, "stream": 0,      "correlation": 1530, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1530, "pid": 2, "tid": 0, "ts": 7429564845664.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845675.011, "dur": 3.607,
    "args": {
      "External id": 4150,      "device": 2, "stream": 0,      "correlation": 1531, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1531, "pid": 2, "tid": 0, "ts": 7429564845675.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845684.392, "dur": 3.407,
    "args": {
      "External id": 4151,      "device": 2, "stream": 0,      "correlation": 1532, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1532, "pid": 2, "tid": 0, "ts": 7429564845684.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845702.514, "dur": 3.768,
    "args": {
      "External id": 4156,      "device": 2, "stream": 0,      "correlation": 1533, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1533, "pid": 2, "tid": 0, "ts": 7429564845702.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564845714.782, "dur": 5.933,
    "args": {
      "External id": 4157,      "device": 2, "stream": 0,      "correlation": 1534, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1534, "pid": 2, "tid": 0, "ts": 7429564845714.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845725.567, "dur": 2.044,
    "args": {
      "External id": 4158,      "device": 2, "stream": 0,      "correlation": 1535, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1535, "pid": 2, "tid": 0, "ts": 7429564845725.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845731.501, "dur": 3.687,
    "args": {
      "External id": 4159,      "device": 2, "stream": 0,      "correlation": 1536, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1536, "pid": 2, "tid": 0, "ts": 7429564845731.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845796.891, "dur": 3.647,
    "args": {
      "External id": 4176,      "device": 2, "stream": 0,      "correlation": 1537, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1537, "pid": 2, "tid": 0, "ts": 7429564845796.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845827.882, "dur": 3.808,
    "args": {
      "External id": 4194,      "device": 2, "stream": 0,      "correlation": 1538, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1538, "pid": 2, "tid": 0, "ts": 7429564845827.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564845881.085, "dur": 5.772,
    "args": {
      "External id": 4207,      "device": 2, "stream": 0,      "correlation": 1542, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1542, "pid": 2, "tid": 0, "ts": 7429564845881.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564845898.565, "dur": 4.248,
    "args": {
      "External id": 4210,      "device": 2, "stream": 0,      "correlation": 1543, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1543, "pid": 2, "tid": 0, "ts": 7429564845898.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564845916.847, "dur": 3.607,
    "args": {
      "External id": 4217,      "device": 2, "stream": 0,      "correlation": 1544, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1544, "pid": 2, "tid": 0, "ts": 7429564845916.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564845935.369, "dur": 3.567,
    "args": {
      "External id": 4220,      "device": 2, "stream": 0,      "correlation": 1545, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1545, "pid": 2, "tid": 0, "ts": 7429564845935.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564845992.020, "dur": 5.812,
    "args": {
      "External id": 4232,      "device": 2, "stream": 0,      "correlation": 1549, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1549, "pid": 2, "tid": 0, "ts": 7429564845992.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846012.106, "dur": 4.168,
    "args": {
      "External id": 4241,      "device": 2, "stream": 0,      "correlation": 1550, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1550, "pid": 2, "tid": 0, "ts": 7429564846012.106,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564846064.867, "dur": 67.875,
    "args": {
      "External id": 4251,      "device": 2, "stream": 0,      "correlation": 1554, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1554, "pid": 2, "tid": 0, "ts": 7429564846064.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564846132.743, "dur": 4.690,
    "args": {
      "External id": 4254,      "device": 2, "stream": 0,      "correlation": 1555, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1555, "pid": 2, "tid": 0, "ts": 7429564846132.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564846137.434, "dur": 4.088,
    "args": {
      "External id": 4256,      "device": 2, "stream": 0,      "correlation": 1556, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1556, "pid": 2, "tid": 0, "ts": 7429564846137.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564846148.219, "dur": 7.857,
    "args": {
      "External id": 4259,      "device": 2, "stream": 0,      "correlation": 1557, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1557, "pid": 2, "tid": 0, "ts": 7429564846148.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564846162.692, "dur": 3.607,
    "args": {
      "External id": 4260,      "device": 2, "stream": 0,      "correlation": 1558, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1558, "pid": 2, "tid": 0, "ts": 7429564846162.692,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564846174.399, "dur": 3.888,
    "args": {
      "External id": 4261,      "device": 2, "stream": 0,      "correlation": 1559, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1559, "pid": 2, "tid": 0, "ts": 7429564846174.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846186.707, "dur": 3.607,
    "args": {
      "External id": 4262,      "device": 2, "stream": 0,      "correlation": 1560, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1560, "pid": 2, "tid": 0, "ts": 7429564846186.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846200.659, "dur": 2.645,
    "args": {
      "External id": 4264,      "device": 2, "stream": 0,      "correlation": 1561, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1561, "pid": 2, "tid": 0, "ts": 7429564846200.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564846260.958, "dur": 76.936,
    "args": {
      "External id": 4272,      "device": 2, "stream": 0,      "correlation": 1565, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1565, "pid": 2, "tid": 0, "ts": 7429564846260.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564846337.895, "dur": 4.569,
    "args": {
      "External id": 4272,      "device": 2, "stream": 0,      "correlation": 1566, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1566, "pid": 2, "tid": 0, "ts": 7429564846337.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564846348.078, "dur": 2.605,
    "args": {
      "External id": 4275,      "device": 2, "stream": 0,      "correlation": 1569, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1569, "pid": 2, "tid": 0, "ts": 7429564846348.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564846396.910, "dur": 81.547,
    "args": {
      "External id": 4283,      "device": 2, "stream": 0,      "correlation": 1573, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1573, "pid": 2, "tid": 0, "ts": 7429564846396.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564846478.458, "dur": 4.890,
    "args": {
      "External id": 4283,      "device": 2, "stream": 0,      "correlation": 1574, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1574, "pid": 2, "tid": 0, "ts": 7429564846478.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564846483.349, "dur": 5.091,
    "args": {
      "External id": 4286,      "device": 2, "stream": 0,      "correlation": 1577, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1577, "pid": 2, "tid": 0, "ts": 7429564846483.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564846488.441, "dur": 86.076,
    "args": {
      "External id": 4294,      "device": 2, "stream": 0,      "correlation": 1581, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1581, "pid": 2, "tid": 0, "ts": 7429564846488.441,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564846574.518, "dur": 4.610,
    "args": {
      "External id": 4294,      "device": 2, "stream": 0,      "correlation": 1582, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1582, "pid": 2, "tid": 0, "ts": 7429564846574.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564846579.129, "dur": 4.690,
    "args": {
      "External id": 4297,      "device": 2, "stream": 0,      "correlation": 1583, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1583, "pid": 2, "tid": 0, "ts": 7429564846579.129,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564846583.820, "dur": 5.050,
    "args": {
      "External id": 4299,      "device": 2, "stream": 0,      "correlation": 1584, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1584, "pid": 2, "tid": 0, "ts": 7429564846583.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564846588.871, "dur": 7.296,
    "args": {
      "External id": 4302,      "device": 2, "stream": 0,      "correlation": 1585, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1585, "pid": 2, "tid": 0, "ts": 7429564846588.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564846596.168, "dur": 4.530,
    "args": {
      "External id": 4303,      "device": 2, "stream": 0,      "correlation": 1586, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1586, "pid": 2, "tid": 0, "ts": 7429564846596.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564846600.699, "dur": 4.569,
    "args": {
      "External id": 4304,      "device": 2, "stream": 0,      "correlation": 1587, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1587, "pid": 2, "tid": 0, "ts": 7429564846600.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846605.269, "dur": 5.291,
    "args": {
      "External id": 4305,      "device": 2, "stream": 0,      "correlation": 1588, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1588, "pid": 2, "tid": 0, "ts": 7429564846605.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846610.561, "dur": 4.129,
    "args": {
      "External id": 4307,      "device": 2, "stream": 0,      "correlation": 1589, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1589, "pid": 2, "tid": 0, "ts": 7429564846610.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564846664.445, "dur": 68.116,
    "args": {
      "External id": 4315,      "device": 2, "stream": 0,      "correlation": 1593, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1593, "pid": 2, "tid": 0, "ts": 7429564846664.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564846732.562, "dur": 16.116,
    "args": {
      "External id": 4328,      "device": 2, "stream": 0,      "correlation": 1597, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1597, "pid": 2, "tid": 0, "ts": 7429564846732.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564846748.679, "dur": 4.088,
    "args": {
      "External id": 4328,      "device": 2, "stream": 0,      "correlation": 1598, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1598, "pid": 2, "tid": 0, "ts": 7429564846748.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564846775.019, "dur": 16.197,
    "args": {
      "External id": 4341,      "device": 2, "stream": 0,      "correlation": 1602, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1602, "pid": 2, "tid": 0, "ts": 7429564846775.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564846791.217, "dur": 4.168,
    "args": {
      "External id": 4341,      "device": 2, "stream": 0,      "correlation": 1603, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1603, "pid": 2, "tid": 0, "ts": 7429564846791.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846804.647, "dur": 3.367,
    "args": {
      "External id": 4351,      "device": 2, "stream": 0,      "correlation": 1604, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1604, "pid": 2, "tid": 0, "ts": 7429564846804.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846833.875, "dur": 2.444,
    "args": {
      "External id": 4356,      "device": 2, "stream": 0,      "correlation": 1605, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1605, "pid": 2, "tid": 0, "ts": 7429564846833.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564846854.763, "dur": 5.852,
    "args": {
      "External id": 4357,      "device": 2, "stream": 0,      "correlation": 1606, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1606, "pid": 2, "tid": 0, "ts": 7429564846854.763,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846867.432, "dur": 2.685,
    "args": {
      "External id": 4358,      "device": 2, "stream": 0,      "correlation": 1607, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1607, "pid": 2, "tid": 0, "ts": 7429564846867.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846877.735, "dur": 3.728,
    "args": {
      "External id": 4359,      "device": 2, "stream": 0,      "correlation": 1608, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1608, "pid": 2, "tid": 0, "ts": 7429564846877.735,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846887.277, "dur": 3.247,
    "args": {
      "External id": 4360,      "device": 2, "stream": 0,      "correlation": 1609, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1609, "pid": 2, "tid": 0, "ts": 7429564846887.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846904.677, "dur": 3.808,
    "args": {
      "External id": 4365,      "device": 2, "stream": 0,      "correlation": 1610, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1610, "pid": 2, "tid": 0, "ts": 7429564846904.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564846915.342, "dur": 6.053,
    "args": {
      "External id": 4366,      "device": 2, "stream": 0,      "correlation": 1611, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1611, "pid": 2, "tid": 0, "ts": 7429564846915.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846927.209, "dur": 3.888,
    "args": {
      "External id": 4367,      "device": 2, "stream": 0,      "correlation": 1612, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1612, "pid": 2, "tid": 0, "ts": 7429564846927.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564846937.032, "dur": 3.687,
    "args": {
      "External id": 4368,      "device": 2, "stream": 0,      "correlation": 1613, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1613, "pid": 2, "tid": 0, "ts": 7429564846937.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564847004.066, "dur": 3.607,
    "args": {
      "External id": 4385,      "device": 2, "stream": 0,      "correlation": 1614, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1614, "pid": 2, "tid": 0, "ts": 7429564847004.066,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564847038.264, "dur": 3.888,
    "args": {
      "External id": 4403,      "device": 2, "stream": 0,      "correlation": 1615, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1615, "pid": 2, "tid": 0, "ts": 7429564847038.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564847089.943, "dur": 5.893,
    "args": {
      "External id": 4416,      "device": 2, "stream": 0,      "correlation": 1619, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1619, "pid": 2, "tid": 0, "ts": 7429564847089.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847106.140, "dur": 3.888,
    "args": {
      "External id": 4419,      "device": 2, "stream": 0,      "correlation": 1620, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1620, "pid": 2, "tid": 0, "ts": 7429564847106.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564847124.944, "dur": 3.687,
    "args": {
      "External id": 4426,      "device": 2, "stream": 0,      "correlation": 1621, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1621, "pid": 2, "tid": 0, "ts": 7429564847124.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564847143.506, "dur": 3.527,
    "args": {
      "External id": 4429,      "device": 2, "stream": 0,      "correlation": 1622, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1622, "pid": 2, "tid": 0, "ts": 7429564847143.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564847199.355, "dur": 5.211,
    "args": {
      "External id": 4441,      "device": 2, "stream": 0,      "correlation": 1626, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1626, "pid": 2, "tid": 0, "ts": 7429564847199.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564847220.002, "dur": 3.968,
    "args": {
      "External id": 4450,      "device": 2, "stream": 0,      "correlation": 1627, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1627, "pid": 2, "tid": 0, "ts": 7429564847220.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564847274.367, "dur": 68.476,
    "args": {
      "External id": 4460,      "device": 2, "stream": 0,      "correlation": 1631, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1631, "pid": 2, "tid": 0, "ts": 7429564847274.367,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847342.844, "dur": 4.650,
    "args": {
      "External id": 4463,      "device": 2, "stream": 0,      "correlation": 1632, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1632, "pid": 2, "tid": 0, "ts": 7429564847342.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847347.495, "dur": 4.890,
    "args": {
      "External id": 4465,      "device": 2, "stream": 0,      "correlation": 1633, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1633, "pid": 2, "tid": 0, "ts": 7429564847347.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564847352.386, "dur": 6.615,
    "args": {
      "External id": 4468,      "device": 2, "stream": 0,      "correlation": 1634, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1634, "pid": 2, "tid": 0, "ts": 7429564847352.386,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847359.002, "dur": 5.892,
    "args": {
      "External id": 4469,      "device": 2, "stream": 0,      "correlation": 1635, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1635, "pid": 2, "tid": 0, "ts": 7429564847359.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847364.895, "dur": 4.730,
    "args": {
      "External id": 4470,      "device": 2, "stream": 0,      "correlation": 1636, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1636, "pid": 2, "tid": 0, "ts": 7429564847364.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564847369.626, "dur": 4.249,
    "args": {
      "External id": 4471,      "device": 2, "stream": 0,      "correlation": 1637, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1637, "pid": 2, "tid": 0, "ts": 7429564847369.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564847434.495, "dur": 4.089,
    "args": {
      "External id": 4473,      "device": 2, "stream": 0,      "correlation": 1640, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1640, "pid": 2, "tid": 0, "ts": 7429564847434.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564847499.645, "dur": 91.409,
    "args": {
      "External id": 4481,      "device": 2, "stream": 0,      "correlation": 1644, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1644, "pid": 2, "tid": 0, "ts": 7429564847499.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564847591.055, "dur": 4.930,
    "args": {
      "External id": 4481,      "device": 2, "stream": 0,      "correlation": 1645, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1645, "pid": 2, "tid": 0, "ts": 7429564847591.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847595.986, "dur": 4.169,
    "args": {
      "External id": 4484,      "device": 2, "stream": 0,      "correlation": 1646, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1646, "pid": 2, "tid": 0, "ts": 7429564847595.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564847606.410, "dur": 82.469,
    "args": {
      "External id": 4492,      "device": 2, "stream": 0,      "correlation": 1652, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1652, "pid": 2, "tid": 0, "ts": 7429564847606.410,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564847688.880, "dur": 4.810,
    "args": {
      "External id": 4492,      "device": 2, "stream": 0,      "correlation": 1653, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1653, "pid": 2, "tid": 0, "ts": 7429564847688.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847693.691, "dur": 4.890,
    "args": {
      "External id": 4495,      "device": 2, "stream": 0,      "correlation": 1654, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1654, "pid": 2, "tid": 0, "ts": 7429564847693.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564847698.582, "dur": 87.280,
    "args": {
      "External id": 4503,      "device": 2, "stream": 0,      "correlation": 1658, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1658, "pid": 2, "tid": 0, "ts": 7429564847698.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564847785.863, "dur": 4.649,
    "args": {
      "External id": 4503,      "device": 2, "stream": 0,      "correlation": 1659, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1659, "pid": 2, "tid": 0, "ts": 7429564847785.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847790.513, "dur": 4.650,
    "args": {
      "External id": 4506,      "device": 2, "stream": 0,      "correlation": 1660, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1660, "pid": 2, "tid": 0, "ts": 7429564847790.513,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847795.164, "dur": 4.690,
    "args": {
      "External id": 4508,      "device": 2, "stream": 0,      "correlation": 1661, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1661, "pid": 2, "tid": 0, "ts": 7429564847795.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564847799.855, "dur": 7.175,
    "args": {
      "External id": 4511,      "device": 2, "stream": 0,      "correlation": 1662, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1662, "pid": 2, "tid": 0, "ts": 7429564847799.855,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847807.031, "dur": 4.490,
    "args": {
      "External id": 4512,      "device": 2, "stream": 0,      "correlation": 1663, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1663, "pid": 2, "tid": 0, "ts": 7429564847807.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564847811.522, "dur": 4.569,
    "args": {
      "External id": 4513,      "device": 2, "stream": 0,      "correlation": 1664, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1664, "pid": 2, "tid": 0, "ts": 7429564847811.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564847816.092, "dur": 4.690,
    "args": {
      "External id": 4514,      "device": 2, "stream": 0,      "correlation": 1665, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1665, "pid": 2, "tid": 0, "ts": 7429564847816.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564847820.783, "dur": 4.129,
    "args": {
      "External id": 4516,      "device": 2, "stream": 0,      "correlation": 1666, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1666, "pid": 2, "tid": 0, "ts": 7429564847820.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564847870.938, "dur": 67.314,
    "args": {
      "External id": 4524,      "device": 2, "stream": 0,      "correlation": 1672, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1672, "pid": 2, "tid": 0, "ts": 7429564847870.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564847938.253, "dur": 15.354,
    "args": {
      "External id": 4537,      "device": 2, "stream": 0,      "correlation": 1676, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1676, "pid": 2, "tid": 0, "ts": 7429564847938.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564847953.608, "dur": 4.089,
    "args": {
      "External id": 4537,      "device": 2, "stream": 0,      "correlation": 1677, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1677, "pid": 2, "tid": 0, "ts": 7429564847953.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564847978.345, "dur": 16.437,
    "args": {
      "External id": 4550,      "device": 2, "stream": 0,      "correlation": 1681, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1681, "pid": 2, "tid": 0, "ts": 7429564847978.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564847994.783, "dur": 3.808,
    "args": {
      "External id": 4550,      "device": 2, "stream": 0,      "correlation": 1682, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1682, "pid": 2, "tid": 0, "ts": 7429564847994.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848012.784, "dur": 2.726,
    "args": {
      "External id": 4560,      "device": 2, "stream": 0,      "correlation": 1683, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1683, "pid": 2, "tid": 0, "ts": 7429564848012.784,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848040.608, "dur": 2.605,
    "args": {
      "External id": 4565,      "device": 2, "stream": 0,      "correlation": 1684, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1684, "pid": 2, "tid": 0, "ts": 7429564848040.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564848059.812, "dur": 6.013,
    "args": {
      "External id": 4566,      "device": 2, "stream": 0,      "correlation": 1685, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1685, "pid": 2, "tid": 0, "ts": 7429564848059.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848073.243, "dur": 3.046,
    "args": {
      "External id": 4567,      "device": 2, "stream": 0,      "correlation": 1686, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1686, "pid": 2, "tid": 0, "ts": 7429564848073.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848083.266, "dur": 2.966,
    "args": {
      "External id": 4568,      "device": 2, "stream": 0,      "correlation": 1687, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1687, "pid": 2, "tid": 0, "ts": 7429564848083.266,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848093.771, "dur": 2.765,
    "args": {
      "External id": 4569,      "device": 2, "stream": 0,      "correlation": 1688, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1688, "pid": 2, "tid": 0, "ts": 7429564848093.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848111.050, "dur": 3.247,
    "args": {
      "External id": 4574,      "device": 2, "stream": 0,      "correlation": 1689, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1689, "pid": 2, "tid": 0, "ts": 7429564848111.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564848124.281, "dur": 5.772,
    "args": {
      "External id": 4575,      "device": 2, "stream": 0,      "correlation": 1690, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1690, "pid": 2, "tid": 0, "ts": 7429564848124.281,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848134.985, "dur": 2.084,
    "args": {
      "External id": 4576,      "device": 2, "stream": 0,      "correlation": 1691, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1691, "pid": 2, "tid": 0, "ts": 7429564848134.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848142.162, "dur": 3.206,
    "args": {
      "External id": 4577,      "device": 2, "stream": 0,      "correlation": 1692, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1692, "pid": 2, "tid": 0, "ts": 7429564848142.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848205.668, "dur": 3.567,
    "args": {
      "External id": 4594,      "device": 2, "stream": 0,      "correlation": 1693, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1693, "pid": 2, "tid": 0, "ts": 7429564848205.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848271.259, "dur": 2.925,
    "args": {
      "External id": 4612,      "device": 2, "stream": 0,      "correlation": 1694, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1694, "pid": 2, "tid": 0, "ts": 7429564848271.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564848320.452, "dur": 5.491,
    "args": {
      "External id": 4625,      "device": 2, "stream": 0,      "correlation": 1698, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1698, "pid": 2, "tid": 0, "ts": 7429564848320.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564848335.486, "dur": 3.287,
    "args": {
      "External id": 4628,      "device": 2, "stream": 0,      "correlation": 1699, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1699, "pid": 2, "tid": 0, "ts": 7429564848335.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848353.167, "dur": 3.086,
    "args": {
      "External id": 4635,      "device": 2, "stream": 0,      "correlation": 1700, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1700, "pid": 2, "tid": 0, "ts": 7429564848353.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564848369.925, "dur": 2.886,
    "args": {
      "External id": 4638,      "device": 2, "stream": 0,      "correlation": 1701, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1701, "pid": 2, "tid": 0, "ts": 7429564848369.925,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564848421.484, "dur": 5.411,
    "args": {
      "External id": 4650,      "device": 2, "stream": 0,      "correlation": 1705, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1705, "pid": 2, "tid": 0, "ts": 7429564848421.484,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848443.093, "dur": 2.926,
    "args": {
      "External id": 4659,      "device": 2, "stream": 0,      "correlation": 1706, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1706, "pid": 2, "tid": 0, "ts": 7429564848443.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564848494.011, "dur": 68.035,
    "args": {
      "External id": 4669,      "device": 2, "stream": 0,      "correlation": 1710, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1710, "pid": 2, "tid": 0, "ts": 7429564848494.011,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564848562.047, "dur": 4.690,
    "args": {
      "External id": 4672,      "device": 2, "stream": 0,      "correlation": 1711, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1711, "pid": 2, "tid": 0, "ts": 7429564848562.047,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564848566.738, "dur": 4.689,
    "args": {
      "External id": 4674,      "device": 2, "stream": 0,      "correlation": 1712, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1712, "pid": 2, "tid": 0, "ts": 7429564848566.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564848571.428, "dur": 7.096,
    "args": {
      "External id": 4677,      "device": 2, "stream": 0,      "correlation": 1713, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1713, "pid": 2, "tid": 0, "ts": 7429564848571.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564848578.525, "dur": 4.449,
    "args": {
      "External id": 4678,      "device": 2, "stream": 0,      "correlation": 1714, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1714, "pid": 2, "tid": 0, "ts": 7429564848578.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564848582.975, "dur": 4.489,
    "args": {
      "External id": 4679,      "device": 2, "stream": 0,      "correlation": 1715, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1715, "pid": 2, "tid": 0, "ts": 7429564848582.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848587.465, "dur": 4.289,
    "args": {
      "External id": 4680,      "device": 2, "stream": 0,      "correlation": 1716, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1716, "pid": 2, "tid": 0, "ts": 7429564848587.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564848595.764, "dur": 3.728,
    "args": {
      "External id": 4682,      "device": 2, "stream": 0,      "correlation": 1717, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1717, "pid": 2, "tid": 0, "ts": 7429564848595.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564848710.067, "dur": 81.426,
    "args": {
      "External id": 4690,      "device": 2, "stream": 0,      "correlation": 1723, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1723, "pid": 2, "tid": 0, "ts": 7429564848710.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564848791.494, "dur": 4.850,
    "args": {
      "External id": 4690,      "device": 2, "stream": 0,      "correlation": 1724, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1724, "pid": 2, "tid": 0, "ts": 7429564848791.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564848796.345, "dur": 4.089,
    "args": {
      "External id": 4693,      "device": 2, "stream": 0,      "correlation": 1725, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1725, "pid": 2, "tid": 0, "ts": 7429564848796.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564848817.113, "dur": 87.160,
    "args": {
      "External id": 4701,      "device": 2, "stream": 0,      "correlation": 1731, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1731, "pid": 2, "tid": 0, "ts": 7429564848817.113,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564848904.274, "dur": 4.770,
    "args": {
      "External id": 4701,      "device": 2, "stream": 0,      "correlation": 1732, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1732, "pid": 2, "tid": 0, "ts": 7429564848904.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564848909.045, "dur": 4.890,
    "args": {
      "External id": 4704,      "device": 2, "stream": 0,      "correlation": 1733, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1733, "pid": 2, "tid": 0, "ts": 7429564848909.045,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564848913.936, "dur": 89.364,
    "args": {
      "External id": 4712,      "device": 2, "stream": 0,      "correlation": 1737, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1737, "pid": 2, "tid": 0, "ts": 7429564848913.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564849003.301, "dur": 4.610,
    "args": {
      "External id": 4712,      "device": 2, "stream": 0,      "correlation": 1738, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1738, "pid": 2, "tid": 0, "ts": 7429564849003.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564849007.912, "dur": 4.730,
    "args": {
      "External id": 4715,      "device": 2, "stream": 0,      "correlation": 1739, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1739, "pid": 2, "tid": 0, "ts": 7429564849007.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564849012.643, "dur": 5.130,
    "args": {
      "External id": 4717,      "device": 2, "stream": 0,      "correlation": 1740, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1740, "pid": 2, "tid": 0, "ts": 7429564849012.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564849017.774, "dur": 7.216,
    "args": {
      "External id": 4720,      "device": 2, "stream": 0,      "correlation": 1741, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1741, "pid": 2, "tid": 0, "ts": 7429564849017.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564849024.991, "dur": 4.529,
    "args": {
      "External id": 4721,      "device": 2, "stream": 0,      "correlation": 1742, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1742, "pid": 2, "tid": 0, "ts": 7429564849024.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564849029.521, "dur": 4.490,
    "args": {
      "External id": 4722,      "device": 2, "stream": 0,      "correlation": 1743, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1743, "pid": 2, "tid": 0, "ts": 7429564849029.521,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849034.012, "dur": 4.690,
    "args": {
      "External id": 4723,      "device": 2, "stream": 0,      "correlation": 1744, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1744, "pid": 2, "tid": 0, "ts": 7429564849034.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849038.703, "dur": 4.088,
    "args": {
      "External id": 4725,      "device": 2, "stream": 0,      "correlation": 1745, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1745, "pid": 2, "tid": 0, "ts": 7429564849038.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564849048.886, "dur": 68.797,
    "args": {
      "External id": 4733,      "device": 2, "stream": 0,      "correlation": 1749, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1749, "pid": 2, "tid": 0, "ts": 7429564849048.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564849117.684, "dur": 16.036,
    "args": {
      "External id": 4746,      "device": 2, "stream": 0,      "correlation": 1753, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1753, "pid": 2, "tid": 0, "ts": 7429564849117.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564849133.721, "dur": 4.008,
    "args": {
      "External id": 4746,      "device": 2, "stream": 0,      "correlation": 1754, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1754, "pid": 2, "tid": 0, "ts": 7429564849133.721,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564849157.055, "dur": 16.998,
    "args": {
      "External id": 4759,      "device": 2, "stream": 0,      "correlation": 1758, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1758, "pid": 2, "tid": 0, "ts": 7429564849157.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564849174.054, "dur": 4.008,
    "args": {
      "External id": 4759,      "device": 2, "stream": 0,      "correlation": 1759, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1759, "pid": 2, "tid": 0, "ts": 7429564849174.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849189.409, "dur": 3.647,
    "args": {
      "External id": 4769,      "device": 2, "stream": 0,      "correlation": 1760, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1760, "pid": 2, "tid": 0, "ts": 7429564849189.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849219.598, "dur": 3.167,
    "args": {
      "External id": 4774,      "device": 2, "stream": 0,      "correlation": 1761, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1761, "pid": 2, "tid": 0, "ts": 7429564849219.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564849241.890, "dur": 6.654,
    "args": {
      "External id": 4775,      "device": 2, "stream": 0,      "correlation": 1762, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1762, "pid": 2, "tid": 0, "ts": 7429564849241.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849295.252, "dur": 3.768,
    "args": {
      "External id": 4776,      "device": 2, "stream": 0,      "correlation": 1765, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1765, "pid": 2, "tid": 0, "ts": 7429564849295.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849307.961, "dur": 3.808,
    "args": {
      "External id": 4777,      "device": 2, "stream": 0,      "correlation": 1766, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1766, "pid": 2, "tid": 0, "ts": 7429564849307.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849319.147, "dur": 3.928,
    "args": {
      "External id": 4778,      "device": 2, "stream": 0,      "correlation": 1767, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1767, "pid": 2, "tid": 0, "ts": 7429564849319.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849337.469, "dur": 4.008,
    "args": {
      "External id": 4783,      "device": 2, "stream": 0,      "correlation": 1768, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1768, "pid": 2, "tid": 0, "ts": 7429564849337.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564849352.183, "dur": 6.213,
    "args": {
      "External id": 4784,      "device": 2, "stream": 0,      "correlation": 1769, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1769, "pid": 2, "tid": 0, "ts": 7429564849352.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849362.006, "dur": 3.407,
    "args": {
      "External id": 4785,      "device": 2, "stream": 0,      "correlation": 1770, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1770, "pid": 2, "tid": 0, "ts": 7429564849362.006,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849370.064, "dur": 3.848,
    "args": {
      "External id": 4786,      "device": 2, "stream": 0,      "correlation": 1771, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1771, "pid": 2, "tid": 0, "ts": 7429564849370.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849439.424, "dur": 3.006,
    "args": {
      "External id": 4803,      "device": 2, "stream": 0,      "correlation": 1772, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1772, "pid": 2, "tid": 0, "ts": 7429564849439.424,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849474.184, "dur": 3.527,
    "args": {
      "External id": 4821,      "device": 2, "stream": 0,      "correlation": 1773, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1773, "pid": 2, "tid": 0, "ts": 7429564849474.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564849526.464, "dur": 5.732,
    "args": {
      "External id": 4834,      "device": 2, "stream": 0,      "correlation": 1777, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1777, "pid": 2, "tid": 0, "ts": 7429564849526.464,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564849543.262, "dur": 3.327,
    "args": {
      "External id": 4837,      "device": 2, "stream": 0,      "correlation": 1778, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1778, "pid": 2, "tid": 0, "ts": 7429564849543.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849561.304, "dur": 3.006,
    "args": {
      "External id": 4844,      "device": 2, "stream": 0,      "correlation": 1779, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1779, "pid": 2, "tid": 0, "ts": 7429564849561.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564849579.105, "dur": 2.765,
    "args": {
      "External id": 4847,      "device": 2, "stream": 0,      "correlation": 1780, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1780, "pid": 2, "tid": 0, "ts": 7429564849579.105,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564849635.915, "dur": 4.730,
    "args": {
      "External id": 4859,      "device": 2, "stream": 0,      "correlation": 1784, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1784, "pid": 2, "tid": 0, "ts": 7429564849635.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849656.483, "dur": 3.046,
    "args": {
      "External id": 4868,      "device": 2, "stream": 0,      "correlation": 1785, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1785, "pid": 2, "tid": 0, "ts": 7429564849656.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564849712.772, "dur": 67.193,
    "args": {
      "External id": 4878,      "device": 2, "stream": 0,      "correlation": 1789, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1789, "pid": 2, "tid": 0, "ts": 7429564849712.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564849779.966, "dur": 4.690,
    "args": {
      "External id": 4881,      "device": 2, "stream": 0,      "correlation": 1790, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1790, "pid": 2, "tid": 0, "ts": 7429564849779.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564849784.657, "dur": 4.890,
    "args": {
      "External id": 4883,      "device": 2, "stream": 0,      "correlation": 1791, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1791, "pid": 2, "tid": 0, "ts": 7429564849784.657,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564849789.548, "dur": 6.975,
    "args": {
      "External id": 4886,      "device": 2, "stream": 0,      "correlation": 1792, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1792, "pid": 2, "tid": 0, "ts": 7429564849789.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564849796.524, "dur": 4.490,
    "args": {
      "External id": 4887,      "device": 2, "stream": 0,      "correlation": 1793, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1793, "pid": 2, "tid": 0, "ts": 7429564849796.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564849801.015, "dur": 4.489,
    "args": {
      "External id": 4888,      "device": 2, "stream": 0,      "correlation": 1794, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1794, "pid": 2, "tid": 0, "ts": 7429564849801.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849805.505, "dur": 4.249,
    "args": {
      "External id": 4889,      "device": 2, "stream": 0,      "correlation": 1795, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1795, "pid": 2, "tid": 0, "ts": 7429564849805.505,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564849815.167, "dur": 3.527,
    "args": {
      "External id": 4891,      "device": 2, "stream": 0,      "correlation": 1796, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1796, "pid": 2, "tid": 0, "ts": 7429564849815.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564849870.254, "dur": 76.495,
    "args": {
      "External id": 4899,      "device": 2, "stream": 0,      "correlation": 1800, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1800, "pid": 2, "tid": 0, "ts": 7429564849870.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564849946.750, "dur": 5.090,
    "args": {
      "External id": 4899,      "device": 2, "stream": 0,      "correlation": 1801, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1801, "pid": 2, "tid": 0, "ts": 7429564849946.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564849951.841, "dur": 4.009,
    "args": {
      "External id": 4902,      "device": 2, "stream": 0,      "correlation": 1804, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1804, "pid": 2, "tid": 0, "ts": 7429564849951.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564849995.823, "dur": 77.938,
    "args": {
      "External id": 4910,      "device": 2, "stream": 0,      "correlation": 1808, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1808, "pid": 2, "tid": 0, "ts": 7429564849995.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564850073.762, "dur": 4.770,
    "args": {
      "External id": 4910,      "device": 2, "stream": 0,      "correlation": 1809, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1809, "pid": 2, "tid": 0, "ts": 7429564850073.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564850078.533, "dur": 4.208,
    "args": {
      "External id": 4913,      "device": 2, "stream": 0,      "correlation": 1812, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1812, "pid": 2, "tid": 0, "ts": 7429564850078.533,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564850105.515, "dur": 87.199,
    "args": {
      "External id": 4921,      "device": 2, "stream": 0,      "correlation": 1816, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1816, "pid": 2, "tid": 0, "ts": 7429564850105.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564850192.715, "dur": 4.650,
    "args": {
      "External id": 4921,      "device": 2, "stream": 0,      "correlation": 1817, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1817, "pid": 2, "tid": 0, "ts": 7429564850192.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564850197.366, "dur": 4.690,
    "args": {
      "External id": 4924,      "device": 2, "stream": 0,      "correlation": 1818, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1818, "pid": 2, "tid": 0, "ts": 7429564850197.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564850202.057, "dur": 4.850,
    "args": {
      "External id": 4926,      "device": 2, "stream": 0,      "correlation": 1819, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1819, "pid": 2, "tid": 0, "ts": 7429564850202.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564850206.908, "dur": 6.855,
    "args": {
      "External id": 4929,      "device": 2, "stream": 0,      "correlation": 1820, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1820, "pid": 2, "tid": 0, "ts": 7429564850206.908,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564850213.764, "dur": 4.529,
    "args": {
      "External id": 4930,      "device": 2, "stream": 0,      "correlation": 1821, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1821, "pid": 2, "tid": 0, "ts": 7429564850213.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564850218.294, "dur": 4.690,
    "args": {
      "External id": 4931,      "device": 2, "stream": 0,      "correlation": 1822, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1822, "pid": 2, "tid": 0, "ts": 7429564850218.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850222.985, "dur": 4.208,
    "args": {
      "External id": 4932,      "device": 2, "stream": 0,      "correlation": 1823, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1823, "pid": 2, "tid": 0, "ts": 7429564850222.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850231.444, "dur": 2.204,
    "args": {
      "External id": 4934,      "device": 2, "stream": 0,      "correlation": 1824, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1824, "pid": 2, "tid": 0, "ts": 7429564850231.444,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564850287.252, "dur": 68.076,
    "args": {
      "External id": 4942,      "device": 2, "stream": 0,      "correlation": 1828, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1828, "pid": 2, "tid": 0, "ts": 7429564850287.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564850355.329, "dur": 15.274,
    "args": {
      "External id": 4955,      "device": 2, "stream": 0,      "correlation": 1832, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1832, "pid": 2, "tid": 0, "ts": 7429564850355.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564850370.604, "dur": 4.008,
    "args": {
      "External id": 4955,      "device": 2, "stream": 0,      "correlation": 1833, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1833, "pid": 2, "tid": 0, "ts": 7429564850370.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564850392.494, "dur": 16.958,
    "args": {
      "External id": 4968,      "device": 2, "stream": 0,      "correlation": 1837, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1837, "pid": 2, "tid": 0, "ts": 7429564850392.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564850409.453, "dur": 4.009,
    "args": {
      "External id": 4968,      "device": 2, "stream": 0,      "correlation": 1838, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1838, "pid": 2, "tid": 0, "ts": 7429564850409.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850422.884, "dur": 3.487,
    "args": {
      "External id": 4978,      "device": 2, "stream": 0,      "correlation": 1839, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1839, "pid": 2, "tid": 0, "ts": 7429564850422.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850450.909, "dur": 3.607,
    "args": {
      "External id": 4983,      "device": 2, "stream": 0,      "correlation": 1840, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1840, "pid": 2, "tid": 0, "ts": 7429564850450.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564850472.719, "dur": 6.694,
    "args": {
      "External id": 4984,      "device": 2, "stream": 0,      "correlation": 1841, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1841, "pid": 2, "tid": 0, "ts": 7429564850472.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850484.867, "dur": 3.487,
    "args": {
      "External id": 4985,      "device": 2, "stream": 0,      "correlation": 1842, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1842, "pid": 2, "tid": 0, "ts": 7429564850484.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850494.850, "dur": 4.128,
    "args": {
      "External id": 4986,      "device": 2, "stream": 0,      "correlation": 1843, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1843, "pid": 2, "tid": 0, "ts": 7429564850494.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850504.432, "dur": 3.687,
    "args": {
      "External id": 4987,      "device": 2, "stream": 0,      "correlation": 1844, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1844, "pid": 2, "tid": 0, "ts": 7429564850504.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850522.112, "dur": 3.688,
    "args": {
      "External id": 4992,      "device": 2, "stream": 0,      "correlation": 1845, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1845, "pid": 2, "tid": 0, "ts": 7429564850522.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564850533.378, "dur": 6.093,
    "args": {
      "External id": 4993,      "device": 2, "stream": 0,      "correlation": 1846, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1846, "pid": 2, "tid": 0, "ts": 7429564850533.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850544.043, "dur": 2.124,
    "args": {
      "External id": 4994,      "device": 2, "stream": 0,      "correlation": 1847, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1847, "pid": 2, "tid": 0, "ts": 7429564850544.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850552.662, "dur": 3.608,
    "args": {
      "External id": 4995,      "device": 2, "stream": 0,      "correlation": 1848, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1848, "pid": 2, "tid": 0, "ts": 7429564850552.662,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850619.697, "dur": 3.607,
    "args": {
      "External id": 5012,      "device": 2, "stream": 0,      "correlation": 1849, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1849, "pid": 2, "tid": 0, "ts": 7429564850619.697,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850654.617, "dur": 3.848,
    "args": {
      "External id": 5030,      "device": 2, "stream": 0,      "correlation": 1850, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1850, "pid": 2, "tid": 0, "ts": 7429564850654.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564850704.772, "dur": 5.933,
    "args": {
      "External id": 5043,      "device": 2, "stream": 0,      "correlation": 1854, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1854, "pid": 2, "tid": 0, "ts": 7429564850704.772,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564850721.611, "dur": 4.890,
    "args": {
      "External id": 5046,      "device": 2, "stream": 0,      "correlation": 1855, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1855, "pid": 2, "tid": 0, "ts": 7429564850721.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850740.534, "dur": 3.688,
    "args": {
      "External id": 5053,      "device": 2, "stream": 0,      "correlation": 1856, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1856, "pid": 2, "tid": 0, "ts": 7429564850740.534,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564850759.578, "dur": 3.688,
    "args": {
      "External id": 5056,      "device": 2, "stream": 0,      "correlation": 1857, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1857, "pid": 2, "tid": 0, "ts": 7429564850759.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564850816.830, "dur": 5.131,
    "args": {
      "External id": 5068,      "device": 2, "stream": 0,      "correlation": 1861, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1861, "pid": 2, "tid": 0, "ts": 7429564850816.830,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850836.114, "dur": 3.808,
    "args": {
      "External id": 5077,      "device": 2, "stream": 0,      "correlation": 1862, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1862, "pid": 2, "tid": 0, "ts": 7429564850836.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564850890.960, "dur": 68.156,
    "args": {
      "External id": 5087,      "device": 2, "stream": 0,      "correlation": 1866, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1866, "pid": 2, "tid": 0, "ts": 7429564850890.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564850959.117, "dur": 4.730,
    "args": {
      "External id": 5090,      "device": 2, "stream": 0,      "correlation": 1867, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1867, "pid": 2, "tid": 0, "ts": 7429564850959.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564850963.848, "dur": 5.491,
    "args": {
      "External id": 5092,      "device": 2, "stream": 0,      "correlation": 1868, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1868, "pid": 2, "tid": 0, "ts": 7429564850963.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564850969.340, "dur": 7.015,
    "args": {
      "External id": 5095,      "device": 2, "stream": 0,      "correlation": 1869, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1869, "pid": 2, "tid": 0, "ts": 7429564850969.340,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564850976.356, "dur": 4.570,
    "args": {
      "External id": 5096,      "device": 2, "stream": 0,      "correlation": 1870, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1870, "pid": 2, "tid": 0, "ts": 7429564850976.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564850980.927, "dur": 4.730,
    "args": {
      "External id": 5097,      "device": 2, "stream": 0,      "correlation": 1871, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1871, "pid": 2, "tid": 0, "ts": 7429564850980.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850985.658, "dur": 5.371,
    "args": {
      "External id": 5098,      "device": 2, "stream": 0,      "correlation": 1872, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1872, "pid": 2, "tid": 0, "ts": 7429564850985.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564850991.030, "dur": 4.209,
    "args": {
      "External id": 5100,      "device": 2, "stream": 0,      "correlation": 1873, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1873, "pid": 2, "tid": 0, "ts": 7429564850991.030,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564851112.469, "dur": 82.830,
    "args": {
      "External id": 5108,      "device": 2, "stream": 0,      "correlation": 1879, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1879, "pid": 2, "tid": 0, "ts": 7429564851112.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564851195.300, "dur": 5.050,
    "args": {
      "External id": 5108,      "device": 2, "stream": 0,      "correlation": 1880, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1880, "pid": 2, "tid": 0, "ts": 7429564851195.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564851200.351, "dur": 4.129,
    "args": {
      "External id": 5111,      "device": 2, "stream": 0,      "correlation": 1881, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1881, "pid": 2, "tid": 0, "ts": 7429564851200.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564851225.088, "dur": 79.582,
    "args": {
      "External id": 5119,      "device": 2, "stream": 0,      "correlation": 1887, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1887, "pid": 2, "tid": 0, "ts": 7429564851225.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564851304.671, "dur": 4.850,
    "args": {
      "External id": 5119,      "device": 2, "stream": 0,      "correlation": 1888, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1888, "pid": 2, "tid": 0, "ts": 7429564851304.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564851309.522, "dur": 4.089,
    "args": {
      "External id": 5122,      "device": 2, "stream": 0,      "correlation": 1889, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1889, "pid": 2, "tid": 0, "ts": 7429564851309.522,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564851317.621, "dur": 86.558,
    "args": {
      "External id": 5130,      "device": 2, "stream": 0,      "correlation": 1893, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1893, "pid": 2, "tid": 0, "ts": 7429564851317.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564851404.180, "dur": 4.770,
    "args": {
      "External id": 5130,      "device": 2, "stream": 0,      "correlation": 1894, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1894, "pid": 2, "tid": 0, "ts": 7429564851404.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564851408.951, "dur": 4.609,
    "args": {
      "External id": 5133,      "device": 2, "stream": 0,      "correlation": 1895, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1895, "pid": 2, "tid": 0, "ts": 7429564851408.951,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564851413.561, "dur": 5.211,
    "args": {
      "External id": 5135,      "device": 2, "stream": 0,      "correlation": 1896, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1896, "pid": 2, "tid": 0, "ts": 7429564851413.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564851418.773, "dur": 7.176,
    "args": {
      "External id": 5138,      "device": 2, "stream": 0,      "correlation": 1897, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1897, "pid": 2, "tid": 0, "ts": 7429564851418.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564851425.950, "dur": 4.409,
    "args": {
      "External id": 5139,      "device": 2, "stream": 0,      "correlation": 1898, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1898, "pid": 2, "tid": 0, "ts": 7429564851425.950,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564851430.360, "dur": 4.610,
    "args": {
      "External id": 5140,      "device": 2, "stream": 0,      "correlation": 1899, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1899, "pid": 2, "tid": 0, "ts": 7429564851430.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851434.971, "dur": 4.689,
    "args": {
      "External id": 5141,      "device": 2, "stream": 0,      "correlation": 1900, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1900, "pid": 2, "tid": 0, "ts": 7429564851434.971,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851439.661, "dur": 4.009,
    "args": {
      "External id": 5143,      "device": 2, "stream": 0,      "correlation": 1901, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1901, "pid": 2, "tid": 0, "ts": 7429564851439.661,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564851487.050, "dur": 67.755,
    "args": {
      "External id": 5151,      "device": 2, "stream": 0,      "correlation": 1905, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1905, "pid": 2, "tid": 0, "ts": 7429564851487.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564851554.806, "dur": 15.595,
    "args": {
      "External id": 5164,      "device": 2, "stream": 0,      "correlation": 1909, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1909, "pid": 2, "tid": 0, "ts": 7429564851554.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564851570.402, "dur": 3.928,
    "args": {
      "External id": 5164,      "device": 2, "stream": 0,      "correlation": 1910, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1910, "pid": 2, "tid": 0, "ts": 7429564851570.402,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564851592.733, "dur": 16.557,
    "args": {
      "External id": 5177,      "device": 2, "stream": 0,      "correlation": 1914, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1914, "pid": 2, "tid": 0, "ts": 7429564851592.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564851609.291, "dur": 4.049,
    "args": {
      "External id": 5177,      "device": 2, "stream": 0,      "correlation": 1915, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1915, "pid": 2, "tid": 0, "ts": 7429564851609.291,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851664.258, "dur": 2.605,
    "args": {
      "External id": 5187,      "device": 2, "stream": 0,      "correlation": 1918, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1918, "pid": 2, "tid": 0, "ts": 7429564851664.258,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851695.770, "dur": 3.046,
    "args": {
      "External id": 5192,      "device": 2, "stream": 0,      "correlation": 1919, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1919, "pid": 2, "tid": 0, "ts": 7429564851695.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564851718.703, "dur": 5.852,
    "args": {
      "External id": 5193,      "device": 2, "stream": 0,      "correlation": 1920, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1920, "pid": 2, "tid": 0, "ts": 7429564851718.703,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851731.733, "dur": 2.885,
    "args": {
      "External id": 5194,      "device": 2, "stream": 0,      "correlation": 1921, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1921, "pid": 2, "tid": 0, "ts": 7429564851731.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851741.996, "dur": 3.407,
    "args": {
      "External id": 5195,      "device": 2, "stream": 0,      "correlation": 1922, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1922, "pid": 2, "tid": 0, "ts": 7429564851741.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851753.503, "dur": 2.324,
    "args": {
      "External id": 5196,      "device": 2, "stream": 0,      "correlation": 1923, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1923, "pid": 2, "tid": 0, "ts": 7429564851753.503,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851770.101, "dur": 2.845,
    "args": {
      "External id": 5201,      "device": 2, "stream": 0,      "correlation": 1924, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1924, "pid": 2, "tid": 0, "ts": 7429564851770.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564851780.164, "dur": 5.411,
    "args": {
      "External id": 5202,      "device": 2, "stream": 0,      "correlation": 1925, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1925, "pid": 2, "tid": 0, "ts": 7429564851780.164,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851790.468, "dur": 2.685,
    "args": {
      "External id": 5203,      "device": 2, "stream": 0,      "correlation": 1926, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1926, "pid": 2, "tid": 0, "ts": 7429564851790.468,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851798.526, "dur": 2.766,
    "args": {
      "External id": 5204,      "device": 2, "stream": 0,      "correlation": 1927, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1927, "pid": 2, "tid": 0, "ts": 7429564851798.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851864.037, "dur": 3.848,
    "args": {
      "External id": 5221,      "device": 2, "stream": 0,      "correlation": 1928, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1928, "pid": 2, "tid": 0, "ts": 7429564851864.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851898.115, "dur": 3.607,
    "args": {
      "External id": 5239,      "device": 2, "stream": 0,      "correlation": 1929, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1929, "pid": 2, "tid": 0, "ts": 7429564851898.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564851948.150, "dur": 5.412,
    "args": {
      "External id": 5252,      "device": 2, "stream": 0,      "correlation": 1933, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1933, "pid": 2, "tid": 0, "ts": 7429564851948.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564851965.310, "dur": 3.727,
    "args": {
      "External id": 5255,      "device": 2, "stream": 0,      "correlation": 1934, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1934, "pid": 2, "tid": 0, "ts": 7429564851965.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564851981.707, "dur": 3.608,
    "args": {
      "External id": 5262,      "device": 2, "stream": 0,      "correlation": 1935, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1935, "pid": 2, "tid": 0, "ts": 7429564851981.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564851998.065, "dur": 3.808,
    "args": {
      "External id": 5265,      "device": 2, "stream": 0,      "correlation": 1936, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1936, "pid": 2, "tid": 0, "ts": 7429564851998.065,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564852050.265, "dur": 5.291,
    "args": {
      "External id": 5277,      "device": 2, "stream": 0,      "correlation": 1940, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1940, "pid": 2, "tid": 0, "ts": 7429564852050.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852070.110, "dur": 3.969,
    "args": {
      "External id": 5286,      "device": 2, "stream": 0,      "correlation": 1941, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1941, "pid": 2, "tid": 0, "ts": 7429564852070.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564852123.794, "dur": 68.196,
    "args": {
      "External id": 5296,      "device": 2, "stream": 0,      "correlation": 1945, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1945, "pid": 2, "tid": 0, "ts": 7429564852123.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564852191.991, "dur": 4.649,
    "args": {
      "External id": 5299,      "device": 2, "stream": 0,      "correlation": 1946, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1946, "pid": 2, "tid": 0, "ts": 7429564852191.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564852196.641, "dur": 4.770,
    "args": {
      "External id": 5301,      "device": 2, "stream": 0,      "correlation": 1947, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1947, "pid": 2, "tid": 0, "ts": 7429564852196.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564852201.412, "dur": 7.216,
    "args": {
      "External id": 5304,      "device": 2, "stream": 0,      "correlation": 1948, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1948, "pid": 2, "tid": 0, "ts": 7429564852201.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564852208.629, "dur": 4.529,
    "args": {
      "External id": 5305,      "device": 2, "stream": 0,      "correlation": 1949, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1949, "pid": 2, "tid": 0, "ts": 7429564852208.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564852213.159, "dur": 4.610,
    "args": {
      "External id": 5306,      "device": 2, "stream": 0,      "correlation": 1950, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1950, "pid": 2, "tid": 0, "ts": 7429564852213.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852217.770, "dur": 4.289,
    "args": {
      "External id": 5307,      "device": 2, "stream": 0,      "correlation": 1951, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1951, "pid": 2, "tid": 0, "ts": 7429564852217.770,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852226.871, "dur": 2.485,
    "args": {
      "External id": 5309,      "device": 2, "stream": 0,      "correlation": 1952, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1952, "pid": 2, "tid": 0, "ts": 7429564852226.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564852279.833, "dur": 82.949,
    "args": {
      "External id": 5317,      "device": 2, "stream": 0,      "correlation": 1956, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1956, "pid": 2, "tid": 0, "ts": 7429564852279.833,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564852362.783, "dur": 4.770,
    "args": {
      "External id": 5317,      "device": 2, "stream": 0,      "correlation": 1957, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1957, "pid": 2, "tid": 0, "ts": 7429564852362.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564852367.554, "dur": 4.089,
    "args": {
      "External id": 5320,      "device": 2, "stream": 0,      "correlation": 1960, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1960, "pid": 2, "tid": 0, "ts": 7429564852367.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564852405.642, "dur": 84.994,
    "args": {
      "External id": 5328,      "device": 2, "stream": 0,      "correlation": 1964, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1964, "pid": 2, "tid": 0, "ts": 7429564852405.642,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564852490.637, "dur": 4.810,
    "args": {
      "External id": 5328,      "device": 2, "stream": 0,      "correlation": 1965, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1965, "pid": 2, "tid": 0, "ts": 7429564852490.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564852495.448, "dur": 5.251,
    "args": {
      "External id": 5331,      "device": 2, "stream": 0,      "correlation": 1968, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1968, "pid": 2, "tid": 0, "ts": 7429564852495.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564852500.700, "dur": 87.400,
    "args": {
      "External id": 5339,      "device": 2, "stream": 0,      "correlation": 1972, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1972, "pid": 2, "tid": 0, "ts": 7429564852500.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564852588.101, "dur": 5.572,
    "args": {
      "External id": 5339,      "device": 2, "stream": 0,      "correlation": 1973, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1973, "pid": 2, "tid": 0, "ts": 7429564852588.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564852593.674, "dur": 4.730,
    "args": {
      "External id": 5342,      "device": 2, "stream": 0,      "correlation": 1974, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1974, "pid": 2, "tid": 0, "ts": 7429564852593.674,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564852598.405, "dur": 4.890,
    "args": {
      "External id": 5344,      "device": 2, "stream": 0,      "correlation": 1975, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1975, "pid": 2, "tid": 0, "ts": 7429564852598.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564852603.296, "dur": 7.176,
    "args": {
      "External id": 5347,      "device": 2, "stream": 0,      "correlation": 1976, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1976, "pid": 2, "tid": 0, "ts": 7429564852603.296,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564852610.473, "dur": 5.451,
    "args": {
      "External id": 5348,      "device": 2, "stream": 0,      "correlation": 1977, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1977, "pid": 2, "tid": 0, "ts": 7429564852610.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564852615.925, "dur": 4.610,
    "args": {
      "External id": 5349,      "device": 2, "stream": 0,      "correlation": 1978, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1978, "pid": 2, "tid": 0, "ts": 7429564852615.925,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852620.536, "dur": 4.649,
    "args": {
      "External id": 5350,      "device": 2, "stream": 0,      "correlation": 1979, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1979, "pid": 2, "tid": 0, "ts": 7429564852620.536,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852625.186, "dur": 4.089,
    "args": {
      "External id": 5352,      "device": 2, "stream": 0,      "correlation": 1980, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1980, "pid": 2, "tid": 0, "ts": 7429564852625.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564852669.608, "dur": 69.239,
    "args": {
      "External id": 5360,      "device": 2, "stream": 0,      "correlation": 1984, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1984, "pid": 2, "tid": 0, "ts": 7429564852669.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564852738.848, "dur": 15.795,
    "args": {
      "External id": 5373,      "device": 2, "stream": 0,      "correlation": 1988, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1988, "pid": 2, "tid": 0, "ts": 7429564852738.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564852754.644, "dur": 4.048,
    "args": {
      "External id": 5373,      "device": 2, "stream": 0,      "correlation": 1989, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1989, "pid": 2, "tid": 0, "ts": 7429564852754.644,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564852774.329, "dur": 16.878,
    "args": {
      "External id": 5386,      "device": 2, "stream": 0,      "correlation": 1993, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1993, "pid": 2, "tid": 0, "ts": 7429564852774.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564852791.208, "dur": 4.128,
    "args": {
      "External id": 5386,      "device": 2, "stream": 0,      "correlation": 1994, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1994, "pid": 2, "tid": 0, "ts": 7429564852791.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852807.405, "dur": 2.685,
    "args": {
      "External id": 5396,      "device": 2, "stream": 0,      "correlation": 1995, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1995, "pid": 2, "tid": 0, "ts": 7429564852807.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852836.512, "dur": 2.445,
    "args": {
      "External id": 5401,      "device": 2, "stream": 0,      "correlation": 1996, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1996, "pid": 2, "tid": 0, "ts": 7429564852836.512,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564852856.318, "dur": 6.133,
    "args": {
      "External id": 5402,      "device": 2, "stream": 0,      "correlation": 1997, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1997, "pid": 2, "tid": 0, "ts": 7429564852856.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852867.944, "dur": 2.485,
    "args": {
      "External id": 5403,      "device": 2, "stream": 0,      "correlation": 1998, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1998, "pid": 2, "tid": 0, "ts": 7429564852867.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852877.246, "dur": 2.765,
    "args": {
      "External id": 5404,      "device": 2, "stream": 0,      "correlation": 1999, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1999, "pid": 2, "tid": 0, "ts": 7429564852877.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852887.710, "dur": 3.046,
    "args": {
      "External id": 5405,      "device": 2, "stream": 0,      "correlation": 2000, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2000, "pid": 2, "tid": 0, "ts": 7429564852887.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852904.027, "dur": 3.327,
    "args": {
      "External id": 5410,      "device": 2, "stream": 0,      "correlation": 2001, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2001, "pid": 2, "tid": 0, "ts": 7429564852904.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564852917.779, "dur": 5.371,
    "args": {
      "External id": 5411,      "device": 2, "stream": 0,      "correlation": 2002, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2002, "pid": 2, "tid": 0, "ts": 7429564852917.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852927.722, "dur": 2.484,
    "args": {
      "External id": 5412,      "device": 2, "stream": 0,      "correlation": 2003, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2003, "pid": 2, "tid": 0, "ts": 7429564852927.722,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564852936.342, "dur": 3.206,
    "args": {
      "External id": 5413,      "device": 2, "stream": 0,      "correlation": 2004, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2004, "pid": 2, "tid": 0, "ts": 7429564852936.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564853001.291, "dur": 2.926,
    "args": {
      "External id": 5430,      "device": 2, "stream": 0,      "correlation": 2005, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2005, "pid": 2, "tid": 0, "ts": 7429564853001.291,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564853034.207, "dur": 3.446,
    "args": {
      "External id": 5448,      "device": 2, "stream": 0,      "correlation": 2006, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2006, "pid": 2, "tid": 0, "ts": 7429564853034.207,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564853084.402, "dur": 5.572,
    "args": {
      "External id": 5461,      "device": 2, "stream": 0,      "correlation": 2010, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2010, "pid": 2, "tid": 0, "ts": 7429564853084.402,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853101.040, "dur": 3.367,
    "args": {
      "External id": 5464,      "device": 2, "stream": 0,      "correlation": 2011, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2011, "pid": 2, "tid": 0, "ts": 7429564853101.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564853118.159, "dur": 2.886,
    "args": {
      "External id": 5471,      "device": 2, "stream": 0,      "correlation": 2012, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2012, "pid": 2, "tid": 0, "ts": 7429564853118.159,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564853135.640, "dur": 3.006,
    "args": {
      "External id": 5474,      "device": 2, "stream": 0,      "correlation": 2013, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2013, "pid": 2, "tid": 0, "ts": 7429564853135.640,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564853189.804, "dur": 4.369,
    "args": {
      "External id": 5486,      "device": 2, "stream": 0,      "correlation": 2017, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2017, "pid": 2, "tid": 0, "ts": 7429564853189.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564853210.051, "dur": 3.126,
    "args": {
      "External id": 5495,      "device": 2, "stream": 0,      "correlation": 2018, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2018, "pid": 2, "tid": 0, "ts": 7429564853210.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564853264.416, "dur": 67.875,
    "args": {
      "External id": 5505,      "device": 2, "stream": 0,      "correlation": 2022, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2022, "pid": 2, "tid": 0, "ts": 7429564853264.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853337.704, "dur": 2.164,
    "args": {
      "External id": 5508,      "device": 2, "stream": 0,      "correlation": 2025, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2025, "pid": 2, "tid": 0, "ts": 7429564853337.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853364.806, "dur": 2.325,
    "args": {
      "External id": 5510,      "device": 2, "stream": 0,      "correlation": 2026, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2026, "pid": 2, "tid": 0, "ts": 7429564853364.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564853381.645, "dur": 7.296,
    "args": {
      "External id": 5513,      "device": 2, "stream": 0,      "correlation": 2027, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2027, "pid": 2, "tid": 0, "ts": 7429564853381.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853394.034, "dur": 3.567,
    "args": {
      "External id": 5514,      "device": 2, "stream": 0,      "correlation": 2028, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2028, "pid": 2, "tid": 0, "ts": 7429564853394.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853404.858, "dur": 3.728,
    "args": {
      "External id": 5515,      "device": 2, "stream": 0,      "correlation": 2029, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2029, "pid": 2, "tid": 0, "ts": 7429564853404.858,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564853418.610, "dur": 3.126,
    "args": {
      "External id": 5516,      "device": 2, "stream": 0,      "correlation": 2030, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2030, "pid": 2, "tid": 0, "ts": 7429564853418.610,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564853431.440, "dur": 2.885,
    "args": {
      "External id": 5518,      "device": 2, "stream": 0,      "correlation": 2031, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2031, "pid": 2, "tid": 0, "ts": 7429564853431.440,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564853491.137, "dur": 81.546,
    "args": {
      "External id": 5526,      "device": 2, "stream": 0,      "correlation": 2035, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2035, "pid": 2, "tid": 0, "ts": 7429564853491.137,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564853572.684, "dur": 4.931,
    "args": {
      "External id": 5526,      "device": 2, "stream": 0,      "correlation": 2036, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2036, "pid": 2, "tid": 0, "ts": 7429564853572.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853577.616, "dur": 4.088,
    "args": {
      "External id": 5529,      "device": 2, "stream": 0,      "correlation": 2037, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2037, "pid": 2, "tid": 0, "ts": 7429564853577.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564853603.114, "dur": 82.870,
    "args": {
      "External id": 5537,      "device": 2, "stream": 0,      "correlation": 2043, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2043, "pid": 2, "tid": 0, "ts": 7429564853603.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564853685.985, "dur": 5.090,
    "args": {
      "External id": 5537,      "device": 2, "stream": 0,      "correlation": 2044, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2044, "pid": 2, "tid": 0, "ts": 7429564853685.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853691.076, "dur": 5.492,
    "args": {
      "External id": 5540,      "device": 2, "stream": 0,      "correlation": 2045, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2045, "pid": 2, "tid": 0, "ts": 7429564853691.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564853696.569, "dur": 88.883,
    "args": {
      "External id": 5548,      "device": 2, "stream": 0,      "correlation": 2049, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2049, "pid": 2, "tid": 0, "ts": 7429564853696.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564853785.453, "dur": 4.610,
    "args": {
      "External id": 5548,      "device": 2, "stream": 0,      "correlation": 2050, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2050, "pid": 2, "tid": 0, "ts": 7429564853785.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853790.064, "dur": 4.770,
    "args": {
      "External id": 5551,      "device": 2, "stream": 0,      "correlation": 2051, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2051, "pid": 2, "tid": 0, "ts": 7429564853790.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853794.835, "dur": 4.650,
    "args": {
      "External id": 5553,      "device": 2, "stream": 0,      "correlation": 2052, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2052, "pid": 2, "tid": 0, "ts": 7429564853794.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564853799.486, "dur": 7.015,
    "args": {
      "External id": 5556,      "device": 2, "stream": 0,      "correlation": 2053, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2053, "pid": 2, "tid": 0, "ts": 7429564853799.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853806.502, "dur": 4.449,
    "args": {
      "External id": 5557,      "device": 2, "stream": 0,      "correlation": 2054, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2054, "pid": 2, "tid": 0, "ts": 7429564853806.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564853810.952, "dur": 6.133,
    "args": {
      "External id": 5558,      "device": 2, "stream": 0,      "correlation": 2055, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2055, "pid": 2, "tid": 0, "ts": 7429564853810.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564853817.086, "dur": 4.650,
    "args": {
      "External id": 5559,      "device": 2, "stream": 0,      "correlation": 2056, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2056, "pid": 2, "tid": 0, "ts": 7429564853817.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564853821.737, "dur": 4.088,
    "args": {
      "External id": 5561,      "device": 2, "stream": 0,      "correlation": 2059, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2059, "pid": 2, "tid": 0, "ts": 7429564853821.737,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564853865.477, "dur": 69.439,
    "args": {
      "External id": 5569,      "device": 2, "stream": 0,      "correlation": 2063, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2063, "pid": 2, "tid": 0, "ts": 7429564853865.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564853934.917, "dur": 15.755,
    "args": {
      "External id": 5582,      "device": 2, "stream": 0,      "correlation": 2067, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2067, "pid": 2, "tid": 0, "ts": 7429564853934.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564853950.673, "dur": 4.209,
    "args": {
      "External id": 5582,      "device": 2, "stream": 0,      "correlation": 2068, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2068, "pid": 2, "tid": 0, "ts": 7429564853950.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564853963.984, "dur": 17.439,
    "args": {
      "External id": 5595,      "device": 2, "stream": 0,      "correlation": 2072, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2072, "pid": 2, "tid": 0, "ts": 7429564853963.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564853981.424, "dur": 3.968,
    "args": {
      "External id": 5595,      "device": 2, "stream": 0,      "correlation": 2073, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2073, "pid": 2, "tid": 0, "ts": 7429564853981.424,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564853995.215, "dur": 3.808,
    "args": {
      "External id": 5605,      "device": 2, "stream": 0,      "correlation": 2074, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2074, "pid": 2, "tid": 0, "ts": 7429564853995.215,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854020.514, "dur": 3.647,
    "args": {
      "External id": 5610,      "device": 2, "stream": 0,      "correlation": 2075, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2075, "pid": 2, "tid": 0, "ts": 7429564854020.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564854040.359, "dur": 6.775,
    "args": {
      "External id": 5611,      "device": 2, "stream": 0,      "correlation": 2076, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2076, "pid": 2, "tid": 0, "ts": 7429564854040.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854053.509, "dur": 3.768,
    "args": {
      "External id": 5612,      "device": 2, "stream": 0,      "correlation": 2077, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2077, "pid": 2, "tid": 0, "ts": 7429564854053.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854063.172, "dur": 3.767,
    "args": {
      "External id": 5613,      "device": 2, "stream": 0,      "correlation": 2078, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2078, "pid": 2, "tid": 0, "ts": 7429564854063.172,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854073.195, "dur": 3.447,
    "args": {
      "External id": 5614,      "device": 2, "stream": 0,      "correlation": 2079, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2079, "pid": 2, "tid": 0, "ts": 7429564854073.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854108.556, "dur": 3.327,
    "args": {
      "External id": 5619,      "device": 2, "stream": 0,      "correlation": 2080, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2080, "pid": 2, "tid": 0, "ts": 7429564854108.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564854130.326, "dur": 6.133,
    "args": {
      "External id": 5620,      "device": 2, "stream": 0,      "correlation": 2081, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2081, "pid": 2, "tid": 0, "ts": 7429564854130.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854143.516, "dur": 3.648,
    "args": {
      "External id": 5621,      "device": 2, "stream": 0,      "correlation": 2082, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2082, "pid": 2, "tid": 0, "ts": 7429564854143.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854154.782, "dur": 3.728,
    "args": {
      "External id": 5622,      "device": 2, "stream": 0,      "correlation": 2083, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2083, "pid": 2, "tid": 0, "ts": 7429564854154.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854249.680, "dur": 4.129,
    "args": {
      "External id": 5639,      "device": 2, "stream": 0,      "correlation": 2084, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2084, "pid": 2, "tid": 0, "ts": 7429564854249.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854281.554, "dur": 3.888,
    "args": {
      "External id": 5657,      "device": 2, "stream": 0,      "correlation": 2085, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2085, "pid": 2, "tid": 0, "ts": 7429564854281.554,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564854338.725, "dur": 6.053,
    "args": {
      "External id": 5670,      "device": 2, "stream": 0,      "correlation": 2089, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2089, "pid": 2, "tid": 0, "ts": 7429564854338.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564854357.448, "dur": 4.489,
    "args": {
      "External id": 5673,      "device": 2, "stream": 0,      "correlation": 2090, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2090, "pid": 2, "tid": 0, "ts": 7429564854357.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854376.131, "dur": 3.768,
    "args": {
      "External id": 5680,      "device": 2, "stream": 0,      "correlation": 2091, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2091, "pid": 2, "tid": 0, "ts": 7429564854376.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564854395.616, "dur": 3.487,
    "args": {
      "External id": 5683,      "device": 2, "stream": 0,      "correlation": 2092, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2092, "pid": 2, "tid": 0, "ts": 7429564854395.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564854451.945, "dur": 5.532,
    "args": {
      "External id": 5695,      "device": 2, "stream": 0,      "correlation": 2096, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2096, "pid": 2, "tid": 0, "ts": 7429564854451.945,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854472.553, "dur": 3.888,
    "args": {
      "External id": 5704,      "device": 2, "stream": 0,      "correlation": 2097, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2097, "pid": 2, "tid": 0, "ts": 7429564854472.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564854530.406, "dur": 68.636,
    "args": {
      "External id": 5714,      "device": 2, "stream": 0,      "correlation": 2101, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2101, "pid": 2, "tid": 0, "ts": 7429564854530.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564854599.043, "dur": 5.773,
    "args": {
      "External id": 5717,      "device": 2, "stream": 0,      "correlation": 2102, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2102, "pid": 2, "tid": 0, "ts": 7429564854599.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564854604.817, "dur": 4.850,
    "args": {
      "External id": 5719,      "device": 2, "stream": 0,      "correlation": 2103, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2103, "pid": 2, "tid": 0, "ts": 7429564854604.817,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564854609.668, "dur": 7.215,
    "args": {
      "External id": 5722,      "device": 2, "stream": 0,      "correlation": 2104, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2104, "pid": 2, "tid": 0, "ts": 7429564854609.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564854616.884, "dur": 4.530,
    "args": {
      "External id": 5723,      "device": 2, "stream": 0,      "correlation": 2105, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2105, "pid": 2, "tid": 0, "ts": 7429564854616.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564854621.415, "dur": 5.411,
    "args": {
      "External id": 5724,      "device": 2, "stream": 0,      "correlation": 2106, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2106, "pid": 2, "tid": 0, "ts": 7429564854621.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854626.827, "dur": 4.249,
    "args": {
      "External id": 5725,      "device": 2, "stream": 0,      "correlation": 2107, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2107, "pid": 2, "tid": 0, "ts": 7429564854626.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564854635.487, "dur": 3.367,
    "args": {
      "External id": 5727,      "device": 2, "stream": 0,      "correlation": 2108, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2108, "pid": 2, "tid": 0, "ts": 7429564854635.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564854759.492, "dur": 78.059,
    "args": {
      "External id": 5735,      "device": 2, "stream": 0,      "correlation": 2114, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2114, "pid": 2, "tid": 0, "ts": 7429564854759.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564854837.552, "dur": 4.730,
    "args": {
      "External id": 5735,      "device": 2, "stream": 0,      "correlation": 2115, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2115, "pid": 2, "tid": 0, "ts": 7429564854837.552,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564854842.283, "dur": 4.168,
    "args": {
      "External id": 5738,      "device": 2, "stream": 0,      "correlation": 2116, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2116, "pid": 2, "tid": 0, "ts": 7429564854842.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564854869.585, "dur": 76.415,
    "args": {
      "External id": 5746,      "device": 2, "stream": 0,      "correlation": 2122, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2122, "pid": 2, "tid": 0, "ts": 7429564854869.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564854946.001, "dur": 4.930,
    "args": {
      "External id": 5746,      "device": 2, "stream": 0,      "correlation": 2123, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2123, "pid": 2, "tid": 0, "ts": 7429564854946.001,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564854950.932, "dur": 4.850,
    "args": {
      "External id": 5749,      "device": 2, "stream": 0,      "correlation": 2124, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2124, "pid": 2, "tid": 0, "ts": 7429564854950.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564854955.783, "dur": 86.639,
    "args": {
      "External id": 5757,      "device": 2, "stream": 0,      "correlation": 2128, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2128, "pid": 2, "tid": 0, "ts": 7429564854955.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564855042.423, "dur": 5.451,
    "args": {
      "External id": 5757,      "device": 2, "stream": 0,      "correlation": 2129, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2129, "pid": 2, "tid": 0, "ts": 7429564855042.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564855047.875, "dur": 4.730,
    "args": {
      "External id": 5760,      "device": 2, "stream": 0,      "correlation": 2130, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2130, "pid": 2, "tid": 0, "ts": 7429564855047.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564855052.606, "dur": 4.690,
    "args": {
      "External id": 5762,      "device": 2, "stream": 0,      "correlation": 2131, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2131, "pid": 2, "tid": 0, "ts": 7429564855052.606,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564855057.297, "dur": 7.095,
    "args": {
      "External id": 5765,      "device": 2, "stream": 0,      "correlation": 2132, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2132, "pid": 2, "tid": 0, "ts": 7429564855057.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564855064.393, "dur": 4.529,
    "args": {
      "External id": 5766,      "device": 2, "stream": 0,      "correlation": 2133, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2133, "pid": 2, "tid": 0, "ts": 7429564855064.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564855068.923, "dur": 4.690,
    "args": {
      "External id": 5767,      "device": 2, "stream": 0,      "correlation": 2134, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2134, "pid": 2, "tid": 0, "ts": 7429564855068.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855073.614, "dur": 4.690,
    "args": {
      "External id": 5768,      "device": 2, "stream": 0,      "correlation": 2135, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2135, "pid": 2, "tid": 0, "ts": 7429564855073.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855078.305, "dur": 4.249,
    "args": {
      "External id": 5770,      "device": 2, "stream": 0,      "correlation": 2136, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2136, "pid": 2, "tid": 0, "ts": 7429564855078.305,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564855097.188, "dur": 67.274,
    "args": {
      "External id": 5778,      "device": 2, "stream": 0,      "correlation": 2140, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2140, "pid": 2, "tid": 0, "ts": 7429564855097.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564855164.463, "dur": 15.635,
    "args": {
      "External id": 5791,      "device": 2, "stream": 0,      "correlation": 2144, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2144, "pid": 2, "tid": 0, "ts": 7429564855164.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564855180.099, "dur": 3.888,
    "args": {
      "External id": 5791,      "device": 2, "stream": 0,      "correlation": 2145, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2145, "pid": 2, "tid": 0, "ts": 7429564855180.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564855202.110, "dur": 16.677,
    "args": {
      "External id": 5804,      "device": 2, "stream": 0,      "correlation": 2149, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2149, "pid": 2, "tid": 0, "ts": 7429564855202.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564855218.788, "dur": 4.008,
    "args": {
      "External id": 5804,      "device": 2, "stream": 0,      "correlation": 2150, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2150, "pid": 2, "tid": 0, "ts": 7429564855218.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855234.624, "dur": 2.686,
    "args": {
      "External id": 5814,      "device": 2, "stream": 0,      "correlation": 2151, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2151, "pid": 2, "tid": 0, "ts": 7429564855234.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855261.246, "dur": 2.404,
    "args": {
      "External id": 5819,      "device": 2, "stream": 0,      "correlation": 2152, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2152, "pid": 2, "tid": 0, "ts": 7429564855261.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564855321.464, "dur": 6.614,
    "args": {
      "External id": 5820,      "device": 2, "stream": 0,      "correlation": 2155, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2155, "pid": 2, "tid": 0, "ts": 7429564855321.464,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855334.093, "dur": 2.846,
    "args": {
      "External id": 5821,      "device": 2, "stream": 0,      "correlation": 2156, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2156, "pid": 2, "tid": 0, "ts": 7429564855334.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855344.316, "dur": 3.648,
    "args": {
      "External id": 5822,      "device": 2, "stream": 0,      "correlation": 2157, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2157, "pid": 2, "tid": 0, "ts": 7429564855344.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855354.580, "dur": 3.968,
    "args": {
      "External id": 5823,      "device": 2, "stream": 0,      "correlation": 2158, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2158, "pid": 2, "tid": 0, "ts": 7429564855354.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855373.664, "dur": 3.768,
    "args": {
      "External id": 5828,      "device": 2, "stream": 0,      "correlation": 2159, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2159, "pid": 2, "tid": 0, "ts": 7429564855373.664,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564855385.491, "dur": 6.334,
    "args": {
      "External id": 5829,      "device": 2, "stream": 0,      "correlation": 2160, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2160, "pid": 2, "tid": 0, "ts": 7429564855385.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855396.396, "dur": 3.407,
    "args": {
      "External id": 5830,      "device": 2, "stream": 0,      "correlation": 2161, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2161, "pid": 2, "tid": 0, "ts": 7429564855396.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855404.375, "dur": 3.887,
    "args": {
      "External id": 5831,      "device": 2, "stream": 0,      "correlation": 2162, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2162, "pid": 2, "tid": 0, "ts": 7429564855404.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855467.600, "dur": 3.888,
    "args": {
      "External id": 5848,      "device": 2, "stream": 0,      "correlation": 2163, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2163, "pid": 2, "tid": 0, "ts": 7429564855467.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855501.678, "dur": 3.928,
    "args": {
      "External id": 5866,      "device": 2, "stream": 0,      "correlation": 2164, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2164, "pid": 2, "tid": 0, "ts": 7429564855501.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564855553.718, "dur": 6.253,
    "args": {
      "External id": 5879,      "device": 2, "stream": 0,      "correlation": 2168, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2168, "pid": 2, "tid": 0, "ts": 7429564855553.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564855570.156, "dur": 4.248,
    "args": {
      "External id": 5882,      "device": 2, "stream": 0,      "correlation": 2169, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2169, "pid": 2, "tid": 0, "ts": 7429564855570.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855588.117, "dur": 3.647,
    "args": {
      "External id": 5889,      "device": 2, "stream": 0,      "correlation": 2170, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2170, "pid": 2, "tid": 0, "ts": 7429564855588.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564855605.316, "dur": 3.407,
    "args": {
      "External id": 5892,      "device": 2, "stream": 0,      "correlation": 2171, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2171, "pid": 2, "tid": 0, "ts": 7429564855605.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564855663.009, "dur": 5.171,
    "args": {
      "External id": 5904,      "device": 2, "stream": 0,      "correlation": 2175, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2175, "pid": 2, "tid": 0, "ts": 7429564855663.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855683.055, "dur": 3.407,
    "args": {
      "External id": 5913,      "device": 2, "stream": 0,      "correlation": 2176, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2176, "pid": 2, "tid": 0, "ts": 7429564855683.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564855737.460, "dur": 68.156,
    "args": {
      "External id": 5923,      "device": 2, "stream": 0,      "correlation": 2180, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2180, "pid": 2, "tid": 0, "ts": 7429564855737.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564855805.617, "dur": 4.770,
    "args": {
      "External id": 5926,      "device": 2, "stream": 0,      "correlation": 2181, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2181, "pid": 2, "tid": 0, "ts": 7429564855805.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564855810.388, "dur": 4.930,
    "args": {
      "External id": 5928,      "device": 2, "stream": 0,      "correlation": 2182, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2182, "pid": 2, "tid": 0, "ts": 7429564855810.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564855815.319, "dur": 6.815,
    "args": {
      "External id": 5931,      "device": 2, "stream": 0,      "correlation": 2183, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2183, "pid": 2, "tid": 0, "ts": 7429564855815.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564855822.135, "dur": 4.569,
    "args": {
      "External id": 5932,      "device": 2, "stream": 0,      "correlation": 2184, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2184, "pid": 2, "tid": 0, "ts": 7429564855822.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564855826.705, "dur": 5.492,
    "args": {
      "External id": 5933,      "device": 2, "stream": 0,      "correlation": 2185, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2185, "pid": 2, "tid": 0, "ts": 7429564855826.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855832.198, "dur": 4.249,
    "args": {
      "External id": 5934,      "device": 2, "stream": 0,      "correlation": 2186, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2186, "pid": 2, "tid": 0, "ts": 7429564855832.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564855842.221, "dur": 3.768,
    "args": {
      "External id": 5936,      "device": 2, "stream": 0,      "correlation": 2187, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2187, "pid": 2, "tid": 0, "ts": 7429564855842.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564855897.588, "dur": 80.905,
    "args": {
      "External id": 5944,      "device": 2, "stream": 0,      "correlation": 2191, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2191, "pid": 2, "tid": 0, "ts": 7429564855897.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564855978.494, "dur": 4.690,
    "args": {
      "External id": 5944,      "device": 2, "stream": 0,      "correlation": 2192, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2192, "pid": 2, "tid": 0, "ts": 7429564855978.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564855983.185, "dur": 3.968,
    "args": {
      "External id": 5947,      "device": 2, "stream": 0,      "correlation": 2195, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2195, "pid": 2, "tid": 0, "ts": 7429564855983.185,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564856024.801, "dur": 80.143,
    "args": {
      "External id": 5955,      "device": 2, "stream": 0,      "correlation": 2199, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2199, "pid": 2, "tid": 0, "ts": 7429564856024.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564856104.945, "dur": 4.770,
    "args": {
      "External id": 5955,      "device": 2, "stream": 0,      "correlation": 2200, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2200, "pid": 2, "tid": 0, "ts": 7429564856104.945,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564856109.716, "dur": 4.208,
    "args": {
      "External id": 5958,      "device": 2, "stream": 0,      "correlation": 2203, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2203, "pid": 2, "tid": 0, "ts": 7429564856109.716,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564856123.588, "dur": 87.600,
    "args": {
      "External id": 5966,      "device": 2, "stream": 0,      "correlation": 2207, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2207, "pid": 2, "tid": 0, "ts": 7429564856123.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564856211.189, "dur": 4.650,
    "args": {
      "External id": 5966,      "device": 2, "stream": 0,      "correlation": 2208, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2208, "pid": 2, "tid": 0, "ts": 7429564856211.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564856215.840, "dur": 4.690,
    "args": {
      "External id": 5969,      "device": 2, "stream": 0,      "correlation": 2209, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2209, "pid": 2, "tid": 0, "ts": 7429564856215.840,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564856220.531, "dur": 4.850,
    "args": {
      "External id": 5971,      "device": 2, "stream": 0,      "correlation": 2210, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2210, "pid": 2, "tid": 0, "ts": 7429564856220.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564856225.382, "dur": 6.814,
    "args": {
      "External id": 5974,      "device": 2, "stream": 0,      "correlation": 2211, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2211, "pid": 2, "tid": 0, "ts": 7429564856225.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564856232.197, "dur": 4.490,
    "args": {
      "External id": 5975,      "device": 2, "stream": 0,      "correlation": 2212, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2212, "pid": 2, "tid": 0, "ts": 7429564856232.197,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564856236.688, "dur": 4.930,
    "args": {
      "External id": 5976,      "device": 2, "stream": 0,      "correlation": 2213, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2213, "pid": 2, "tid": 0, "ts": 7429564856236.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856241.619, "dur": 4.730,
    "args": {
      "External id": 5977,      "device": 2, "stream": 0,      "correlation": 2214, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2214, "pid": 2, "tid": 0, "ts": 7429564856241.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856246.350, "dur": 4.088,
    "args": {
      "External id": 5979,      "device": 2, "stream": 0,      "correlation": 2215, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2215, "pid": 2, "tid": 0, "ts": 7429564856246.350,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564856295.824, "dur": 68.115,
    "args": {
      "External id": 5987,      "device": 2, "stream": 0,      "correlation": 2219, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2219, "pid": 2, "tid": 0, "ts": 7429564856295.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564856363.940, "dur": 15.515,
    "args": {
      "External id": 6000,      "device": 2, "stream": 0,      "correlation": 2223, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2223, "pid": 2, "tid": 0, "ts": 7429564856363.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564856379.456, "dur": 4.088,
    "args": {
      "External id": 6000,      "device": 2, "stream": 0,      "correlation": 2224, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2224, "pid": 2, "tid": 0, "ts": 7429564856379.456,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564856397.337, "dur": 16.838,
    "args": {
      "External id": 6013,      "device": 2, "stream": 0,      "correlation": 2228, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2228, "pid": 2, "tid": 0, "ts": 7429564856397.337,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564856414.176, "dur": 4.048,
    "args": {
      "External id": 6013,      "device": 2, "stream": 0,      "correlation": 2229, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2229, "pid": 2, "tid": 0, "ts": 7429564856414.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856430.052, "dur": 2.966,
    "args": {
      "External id": 6023,      "device": 2, "stream": 0,      "correlation": 2230, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2230, "pid": 2, "tid": 0, "ts": 7429564856430.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856459.319, "dur": 3.127,
    "args": {
      "External id": 6028,      "device": 2, "stream": 0,      "correlation": 2231, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2231, "pid": 2, "tid": 0, "ts": 7429564856459.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564856479.446, "dur": 6.694,
    "args": {
      "External id": 6029,      "device": 2, "stream": 0,      "correlation": 2232, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2232, "pid": 2, "tid": 0, "ts": 7429564856479.446,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856492.556, "dur": 2.565,
    "args": {
      "External id": 6030,      "device": 2, "stream": 0,      "correlation": 2233, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2233, "pid": 2, "tid": 0, "ts": 7429564856492.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856502.178, "dur": 3.206,
    "args": {
      "External id": 6031,      "device": 2, "stream": 0,      "correlation": 2234, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2234, "pid": 2, "tid": 0, "ts": 7429564856502.178,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856512.442, "dur": 3.246,
    "args": {
      "External id": 6032,      "device": 2, "stream": 0,      "correlation": 2235, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2235, "pid": 2, "tid": 0, "ts": 7429564856512.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856529.641, "dur": 3.086,
    "args": {
      "External id": 6037,      "device": 2, "stream": 0,      "correlation": 2236, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2236, "pid": 2, "tid": 0, "ts": 7429564856529.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564856542.751, "dur": 5.412,
    "args": {
      "External id": 6038,      "device": 2, "stream": 0,      "correlation": 2237, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2237, "pid": 2, "tid": 0, "ts": 7429564856542.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856552.975, "dur": 2.164,
    "args": {
      "External id": 6039,      "device": 2, "stream": 0,      "correlation": 2238, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2238, "pid": 2, "tid": 0, "ts": 7429564856552.975,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856559.790, "dur": 3.968,
    "args": {
      "External id": 6040,      "device": 2, "stream": 0,      "correlation": 2239, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2239, "pid": 2, "tid": 0, "ts": 7429564856559.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856625.662, "dur": 3.006,
    "args": {
      "External id": 6057,      "device": 2, "stream": 0,      "correlation": 2240, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2240, "pid": 2, "tid": 0, "ts": 7429564856625.662,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856660.863, "dur": 3.727,
    "args": {
      "External id": 6075,      "device": 2, "stream": 0,      "correlation": 2241, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2241, "pid": 2, "tid": 0, "ts": 7429564856660.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564856710.617, "dur": 5.211,
    "args": {
      "External id": 6088,      "device": 2, "stream": 0,      "correlation": 2245, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2245, "pid": 2, "tid": 0, "ts": 7429564856710.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564856728.618, "dur": 3.528,
    "args": {
      "External id": 6091,      "device": 2, "stream": 0,      "correlation": 2246, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2246, "pid": 2, "tid": 0, "ts": 7429564856728.618,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856746.660, "dur": 2.485,
    "args": {
      "External id": 6098,      "device": 2, "stream": 0,      "correlation": 2247, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2247, "pid": 2, "tid": 0, "ts": 7429564856746.660,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564856762.617, "dur": 3.286,
    "args": {
      "External id": 6101,      "device": 2, "stream": 0,      "correlation": 2248, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2248, "pid": 2, "tid": 0, "ts": 7429564856762.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564856818.866, "dur": 5.091,
    "args": {
      "External id": 6113,      "device": 2, "stream": 0,      "correlation": 2252, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2252, "pid": 2, "tid": 0, "ts": 7429564856818.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856838.711, "dur": 3.728,
    "args": {
      "External id": 6122,      "device": 2, "stream": 0,      "correlation": 2253, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2253, "pid": 2, "tid": 0, "ts": 7429564856838.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564856889.388, "dur": 68.757,
    "args": {
      "External id": 6132,      "device": 2, "stream": 0,      "correlation": 2257, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2257, "pid": 2, "tid": 0, "ts": 7429564856889.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564856958.146, "dur": 4.850,
    "args": {
      "External id": 6135,      "device": 2, "stream": 0,      "correlation": 2258, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2258, "pid": 2, "tid": 0, "ts": 7429564856958.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564856962.997, "dur": 5.893,
    "args": {
      "External id": 6137,      "device": 2, "stream": 0,      "correlation": 2259, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2259, "pid": 2, "tid": 0, "ts": 7429564856962.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564856968.891, "dur": 6.854,
    "args": {
      "External id": 6140,      "device": 2, "stream": 0,      "correlation": 2260, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2260, "pid": 2, "tid": 0, "ts": 7429564856968.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564856975.746, "dur": 4.530,
    "args": {
      "External id": 6141,      "device": 2, "stream": 0,      "correlation": 2261, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2261, "pid": 2, "tid": 0, "ts": 7429564856975.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564856980.277, "dur": 4.690,
    "args": {
      "External id": 6142,      "device": 2, "stream": 0,      "correlation": 2262, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2262, "pid": 2, "tid": 0, "ts": 7429564856980.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564856984.968, "dur": 4.369,
    "args": {
      "External id": 6143,      "device": 2, "stream": 0,      "correlation": 2263, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2263, "pid": 2, "tid": 0, "ts": 7429564856984.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857075.255, "dur": 4.289,
    "args": {
      "External id": 6145,      "device": 2, "stream": 0,      "correlation": 2266, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2266, "pid": 2, "tid": 0, "ts": 7429564857075.255,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564857137.639, "dur": 85.555,
    "args": {
      "External id": 6153,      "device": 2, "stream": 0,      "correlation": 2270, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2270, "pid": 2, "tid": 0, "ts": 7429564857137.639,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564857223.195, "dur": 4.850,
    "args": {
      "External id": 6153,      "device": 2, "stream": 0,      "correlation": 2271, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2271, "pid": 2, "tid": 0, "ts": 7429564857223.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564857228.046, "dur": 4.089,
    "args": {
      "External id": 6156,      "device": 2, "stream": 0,      "correlation": 2272, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2272, "pid": 2, "tid": 0, "ts": 7429564857228.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564857242.520, "dur": 81.506,
    "args": {
      "External id": 6164,      "device": 2, "stream": 0,      "correlation": 2278, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2278, "pid": 2, "tid": 0, "ts": 7429564857242.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564857324.027, "dur": 5.091,
    "args": {
      "External id": 6164,      "device": 2, "stream": 0,      "correlation": 2279, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2279, "pid": 2, "tid": 0, "ts": 7429564857324.027,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564857329.119, "dur": 4.930,
    "args": {
      "External id": 6167,      "device": 2, "stream": 0,      "correlation": 2280, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2280, "pid": 2, "tid": 0, "ts": 7429564857329.119,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564857334.050, "dur": 86.358,
    "args": {
      "External id": 6175,      "device": 2, "stream": 0,      "correlation": 2284, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2284, "pid": 2, "tid": 0, "ts": 7429564857334.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564857420.409, "dur": 4.609,
    "args": {
      "External id": 6175,      "device": 2, "stream": 0,      "correlation": 2285, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2285, "pid": 2, "tid": 0, "ts": 7429564857420.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564857425.019, "dur": 4.690,
    "args": {
      "External id": 6178,      "device": 2, "stream": 0,      "correlation": 2286, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2286, "pid": 2, "tid": 0, "ts": 7429564857425.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564857429.710, "dur": 4.690,
    "args": {
      "External id": 6180,      "device": 2, "stream": 0,      "correlation": 2287, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2287, "pid": 2, "tid": 0, "ts": 7429564857429.710,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564857434.401, "dur": 7.336,
    "args": {
      "External id": 6183,      "device": 2, "stream": 0,      "correlation": 2288, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2288, "pid": 2, "tid": 0, "ts": 7429564857434.401,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564857441.738, "dur": 4.449,
    "args": {
      "External id": 6184,      "device": 2, "stream": 0,      "correlation": 2289, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2289, "pid": 2, "tid": 0, "ts": 7429564857441.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564857446.188, "dur": 4.529,
    "args": {
      "External id": 6185,      "device": 2, "stream": 0,      "correlation": 2290, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2290, "pid": 2, "tid": 0, "ts": 7429564857446.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857450.718, "dur": 4.690,
    "args": {
      "External id": 6186,      "device": 2, "stream": 0,      "correlation": 2291, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2291, "pid": 2, "tid": 0, "ts": 7429564857450.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857455.409, "dur": 4.249,
    "args": {
      "External id": 6188,      "device": 2, "stream": 0,      "correlation": 2292, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2292, "pid": 2, "tid": 0, "ts": 7429564857455.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564857503.199, "dur": 68.757,
    "args": {
      "External id": 6196,      "device": 2, "stream": 0,      "correlation": 2298, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2298, "pid": 2, "tid": 0, "ts": 7429564857503.199,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564857571.957, "dur": 15.474,
    "args": {
      "External id": 6209,      "device": 2, "stream": 0,      "correlation": 2302, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2302, "pid": 2, "tid": 0, "ts": 7429564857571.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564857587.432, "dur": 3.969,
    "args": {
      "External id": 6209,      "device": 2, "stream": 0,      "correlation": 2303, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2303, "pid": 2, "tid": 0, "ts": 7429564857587.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564857613.332, "dur": 17.078,
    "args": {
      "External id": 6222,      "device": 2, "stream": 0,      "correlation": 2307, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2307, "pid": 2, "tid": 0, "ts": 7429564857613.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564857630.411, "dur": 4.169,
    "args": {
      "External id": 6222,      "device": 2, "stream": 0,      "correlation": 2308, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2308, "pid": 2, "tid": 0, "ts": 7429564857630.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857645.566, "dur": 3.688,
    "args": {
      "External id": 6232,      "device": 2, "stream": 0,      "correlation": 2309, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2309, "pid": 2, "tid": 0, "ts": 7429564857645.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857674.072, "dur": 3.406,
    "args": {
      "External id": 6237,      "device": 2, "stream": 0,      "correlation": 2310, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2310, "pid": 2, "tid": 0, "ts": 7429564857674.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564857696.042, "dur": 7.015,
    "args": {
      "External id": 6238,      "device": 2, "stream": 0,      "correlation": 2311, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2311, "pid": 2, "tid": 0, "ts": 7429564857696.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857708.791, "dur": 3.528,
    "args": {
      "External id": 6239,      "device": 2, "stream": 0,      "correlation": 2312, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2312, "pid": 2, "tid": 0, "ts": 7429564857708.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857718.734, "dur": 3.848,
    "args": {
      "External id": 6240,      "device": 2, "stream": 0,      "correlation": 2313, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2313, "pid": 2, "tid": 0, "ts": 7429564857718.734,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857728.356, "dur": 3.688,
    "args": {
      "External id": 6241,      "device": 2, "stream": 0,      "correlation": 2314, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2314, "pid": 2, "tid": 0, "ts": 7429564857728.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857745.516, "dur": 3.808,
    "args": {
      "External id": 6246,      "device": 2, "stream": 0,      "correlation": 2315, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2315, "pid": 2, "tid": 0, "ts": 7429564857745.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564857758.345, "dur": 6.053,
    "args": {
      "External id": 6247,      "device": 2, "stream": 0,      "correlation": 2316, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2316, "pid": 2, "tid": 0, "ts": 7429564857758.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857768.088, "dur": 3.246,
    "args": {
      "External id": 6248,      "device": 2, "stream": 0,      "correlation": 2317, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2317, "pid": 2, "tid": 0, "ts": 7429564857768.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857776.667, "dur": 3.487,
    "args": {
      "External id": 6249,      "device": 2, "stream": 0,      "correlation": 2318, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2318, "pid": 2, "tid": 0, "ts": 7429564857776.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857843.301, "dur": 2.845,
    "args": {
      "External id": 6266,      "device": 2, "stream": 0,      "correlation": 2319, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2319, "pid": 2, "tid": 0, "ts": 7429564857843.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857877.178, "dur": 3.407,
    "args": {
      "External id": 6284,      "device": 2, "stream": 0,      "correlation": 2320, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2320, "pid": 2, "tid": 0, "ts": 7429564857877.178,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564857926.412, "dur": 5.491,
    "args": {
      "External id": 6297,      "device": 2, "stream": 0,      "correlation": 2324, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2324, "pid": 2, "tid": 0, "ts": 7429564857926.412,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564857943.531, "dur": 3.487,
    "args": {
      "External id": 6300,      "device": 2, "stream": 0,      "correlation": 2325, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2325, "pid": 2, "tid": 0, "ts": 7429564857943.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564857961.733, "dur": 2.525,
    "args": {
      "External id": 6307,      "device": 2, "stream": 0,      "correlation": 2326, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2326, "pid": 2, "tid": 0, "ts": 7429564857961.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564857979.895, "dur": 3.126,
    "args": {
      "External id": 6310,      "device": 2, "stream": 0,      "correlation": 2327, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2327, "pid": 2, "tid": 0, "ts": 7429564857979.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564858037.748, "dur": 4.529,
    "args": {
      "External id": 6322,      "device": 2, "stream": 0,      "correlation": 2331, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2331, "pid": 2, "tid": 0, "ts": 7429564858037.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858058.836, "dur": 2.645,
    "args": {
      "External id": 6331,      "device": 2, "stream": 0,      "correlation": 2332, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2332, "pid": 2, "tid": 0, "ts": 7429564858058.836,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564858144.232, "dur": 67.194,
    "args": {
      "External id": 6341,      "device": 2, "stream": 0,      "correlation": 2336, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2336, "pid": 2, "tid": 0, "ts": 7429564858144.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564858211.427, "dur": 4.609,
    "args": {
      "External id": 6344,      "device": 2, "stream": 0,      "correlation": 2337, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2337, "pid": 2, "tid": 0, "ts": 7429564858211.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564858216.037, "dur": 4.209,
    "args": {
      "External id": 6346,      "device": 2, "stream": 0,      "correlation": 2338, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2338, "pid": 2, "tid": 0, "ts": 7429564858216.037,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564858226.381, "dur": 6.614,
    "args": {
      "External id": 6349,      "device": 2, "stream": 0,      "correlation": 2339, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2339, "pid": 2, "tid": 0, "ts": 7429564858226.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564858232.996, "dur": 4.530,
    "args": {
      "External id": 6350,      "device": 2, "stream": 0,      "correlation": 2340, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2340, "pid": 2, "tid": 0, "ts": 7429564858232.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564858237.527, "dur": 4.168,
    "args": {
      "External id": 6351,      "device": 2, "stream": 0,      "correlation": 2341, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2341, "pid": 2, "tid": 0, "ts": 7429564858237.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858246.347, "dur": 3.367,
    "args": {
      "External id": 6352,      "device": 2, "stream": 0,      "correlation": 2342, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2342, "pid": 2, "tid": 0, "ts": 7429564858246.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858260.860, "dur": 2.726,
    "args": {
      "External id": 6354,      "device": 2, "stream": 0,      "correlation": 2343, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2343, "pid": 2, "tid": 0, "ts": 7429564858260.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564858383.302, "dur": 85.596,
    "args": {
      "External id": 6362,      "device": 2, "stream": 0,      "correlation": 2349, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2349, "pid": 2, "tid": 0, "ts": 7429564858383.302,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564858468.899, "dur": 4.970,
    "args": {
      "External id": 6362,      "device": 2, "stream": 0,      "correlation": 2350, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2350, "pid": 2, "tid": 0, "ts": 7429564858468.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564858473.870, "dur": 4.129,
    "args": {
      "External id": 6365,      "device": 2, "stream": 0,      "correlation": 2351, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2351, "pid": 2, "tid": 0, "ts": 7429564858473.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564858493.194, "dur": 84.233,
    "args": {
      "External id": 6373,      "device": 2, "stream": 0,      "correlation": 2357, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2357, "pid": 2, "tid": 0, "ts": 7429564858493.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564858577.428, "dur": 4.930,
    "args": {
      "External id": 6373,      "device": 2, "stream": 0,      "correlation": 2358, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2358, "pid": 2, "tid": 0, "ts": 7429564858577.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564858582.359, "dur": 5.011,
    "args": {
      "External id": 6376,      "device": 2, "stream": 0,      "correlation": 2359, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2359, "pid": 2, "tid": 0, "ts": 7429564858582.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564858587.371, "dur": 89.605,
    "args": {
      "External id": 6384,      "device": 2, "stream": 0,      "correlation": 2363, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2363, "pid": 2, "tid": 0, "ts": 7429564858587.371,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564858676.977, "dur": 4.489,
    "args": {
      "External id": 6384,      "device": 2, "stream": 0,      "correlation": 2364, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2364, "pid": 2, "tid": 0, "ts": 7429564858676.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564858681.467, "dur": 4.650,
    "args": {
      "External id": 6387,      "device": 2, "stream": 0,      "correlation": 2365, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2365, "pid": 2, "tid": 0, "ts": 7429564858681.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564858686.118, "dur": 4.569,
    "args": {
      "External id": 6389,      "device": 2, "stream": 0,      "correlation": 2366, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2366, "pid": 2, "tid": 0, "ts": 7429564858686.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564858690.688, "dur": 7.056,
    "args": {
      "External id": 6392,      "device": 2, "stream": 0,      "correlation": 2367, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2367, "pid": 2, "tid": 0, "ts": 7429564858690.688,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564858697.745, "dur": 4.529,
    "args": {
      "External id": 6393,      "device": 2, "stream": 0,      "correlation": 2368, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2368, "pid": 2, "tid": 0, "ts": 7429564858697.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564858702.275, "dur": 4.690,
    "args": {
      "External id": 6394,      "device": 2, "stream": 0,      "correlation": 2369, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2369, "pid": 2, "tid": 0, "ts": 7429564858702.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858706.966, "dur": 4.730,
    "args": {
      "External id": 6395,      "device": 2, "stream": 0,      "correlation": 2370, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2370, "pid": 2, "tid": 0, "ts": 7429564858706.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858711.697, "dur": 4.048,
    "args": {
      "External id": 6397,      "device": 2, "stream": 0,      "correlation": 2371, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2371, "pid": 2, "tid": 0, "ts": 7429564858711.697,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564858720.798, "dur": 67.153,
    "args": {
      "External id": 6405,      "device": 2, "stream": 0,      "correlation": 2375, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2375, "pid": 2, "tid": 0, "ts": 7429564858720.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564858787.952, "dur": 15.555,
    "args": {
      "External id": 6418,      "device": 2, "stream": 0,      "correlation": 2379, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2379, "pid": 2, "tid": 0, "ts": 7429564858787.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564858803.508, "dur": 4.088,
    "args": {
      "External id": 6418,      "device": 2, "stream": 0,      "correlation": 2380, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2380, "pid": 2, "tid": 0, "ts": 7429564858803.508,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564858821.028, "dur": 17.279,
    "args": {
      "External id": 6431,      "device": 2, "stream": 0,      "correlation": 2384, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2384, "pid": 2, "tid": 0, "ts": 7429564858821.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564858838.308, "dur": 4.088,
    "args": {
      "External id": 6431,      "device": 2, "stream": 0,      "correlation": 2385, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2385, "pid": 2, "tid": 0, "ts": 7429564858838.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858852.981, "dur": 3.688,
    "args": {
      "External id": 6441,      "device": 2, "stream": 0,      "correlation": 2386, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2386, "pid": 2, "tid": 0, "ts": 7429564858852.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858880.805, "dur": 3.487,
    "args": {
      "External id": 6446,      "device": 2, "stream": 0,      "correlation": 2387, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2387, "pid": 2, "tid": 0, "ts": 7429564858880.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564858903.698, "dur": 6.694,
    "args": {
      "External id": 6447,      "device": 2, "stream": 0,      "correlation": 2388, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2388, "pid": 2, "tid": 0, "ts": 7429564858903.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858956.259, "dur": 3.928,
    "args": {
      "External id": 6448,      "device": 2, "stream": 0,      "correlation": 2391, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2391, "pid": 2, "tid": 0, "ts": 7429564858956.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858969.329, "dur": 3.768,
    "args": {
      "External id": 6449,      "device": 2, "stream": 0,      "correlation": 2392, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2392, "pid": 2, "tid": 0, "ts": 7429564858969.329,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858979.392, "dur": 3.768,
    "args": {
      "External id": 6450,      "device": 2, "stream": 0,      "correlation": 2393, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2393, "pid": 2, "tid": 0, "ts": 7429564858979.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564858999.759, "dur": 3.607,
    "args": {
      "External id": 6455,      "device": 2, "stream": 0,      "correlation": 2394, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2394, "pid": 2, "tid": 0, "ts": 7429564858999.759,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564859012.308, "dur": 6.333,
    "args": {
      "External id": 6456,      "device": 2, "stream": 0,      "correlation": 2395, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2395, "pid": 2, "tid": 0, "ts": 7429564859012.308,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564859022.651, "dur": 3.087,
    "args": {
      "External id": 6457,      "device": 2, "stream": 0,      "correlation": 2396, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2396, "pid": 2, "tid": 0, "ts": 7429564859022.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564859031.191, "dur": 3.688,
    "args": {
      "External id": 6458,      "device": 2, "stream": 0,      "correlation": 2397, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2397, "pid": 2, "tid": 0, "ts": 7429564859031.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564859097.463, "dur": 3.648,
    "args": {
      "External id": 6475,      "device": 2, "stream": 0,      "correlation": 2398, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2398, "pid": 2, "tid": 0, "ts": 7429564859097.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564859132.063, "dur": 3.808,
    "args": {
      "External id": 6493,      "device": 2, "stream": 0,      "correlation": 2399, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2399, "pid": 2, "tid": 0, "ts": 7429564859132.063,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564859183.100, "dur": 6.254,
    "args": {
      "External id": 6506,      "device": 2, "stream": 0,      "correlation": 2403, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2403, "pid": 2, "tid": 0, "ts": 7429564859183.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859199.538, "dur": 4.048,
    "args": {
      "External id": 6509,      "device": 2, "stream": 0,      "correlation": 2404, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2404, "pid": 2, "tid": 0, "ts": 7429564859199.538,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564859217.219, "dur": 3.647,
    "args": {
      "External id": 6516,      "device": 2, "stream": 0,      "correlation": 2405, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2405, "pid": 2, "tid": 0, "ts": 7429564859217.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564859235.060, "dur": 3.447,
    "args": {
      "External id": 6519,      "device": 2, "stream": 0,      "correlation": 2406, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2406, "pid": 2, "tid": 0, "ts": 7429564859235.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564859291.068, "dur": 5.452,
    "args": {
      "External id": 6531,      "device": 2, "stream": 0,      "correlation": 2410, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2410, "pid": 2, "tid": 0, "ts": 7429564859291.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564859310.112, "dur": 3.688,
    "args": {
      "External id": 6540,      "device": 2, "stream": 0,      "correlation": 2411, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2411, "pid": 2, "tid": 0, "ts": 7429564859310.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564859364.718, "dur": 69.118,
    "args": {
      "External id": 6550,      "device": 2, "stream": 0,      "correlation": 2415, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2415, "pid": 2, "tid": 0, "ts": 7429564859364.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859433.837, "dur": 4.729,
    "args": {
      "External id": 6553,      "device": 2, "stream": 0,      "correlation": 2416, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2416, "pid": 2, "tid": 0, "ts": 7429564859433.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859438.567, "dur": 4.971,
    "args": {
      "External id": 6555,      "device": 2, "stream": 0,      "correlation": 2417, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2417, "pid": 2, "tid": 0, "ts": 7429564859438.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564859443.539, "dur": 6.895,
    "args": {
      "External id": 6558,      "device": 2, "stream": 0,      "correlation": 2418, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2418, "pid": 2, "tid": 0, "ts": 7429564859443.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859450.435, "dur": 4.529,
    "args": {
      "External id": 6559,      "device": 2, "stream": 0,      "correlation": 2419, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2419, "pid": 2, "tid": 0, "ts": 7429564859450.435,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859454.965, "dur": 5.011,
    "args": {
      "External id": 6560,      "device": 2, "stream": 0,      "correlation": 2420, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2420, "pid": 2, "tid": 0, "ts": 7429564859454.965,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564859459.977, "dur": 4.168,
    "args": {
      "External id": 6561,      "device": 2, "stream": 0,      "correlation": 2421, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2421, "pid": 2, "tid": 0, "ts": 7429564859459.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564859470.962, "dur": 2.605,
    "args": {
      "External id": 6563,      "device": 2, "stream": 0,      "correlation": 2422, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2422, "pid": 2, "tid": 0, "ts": 7429564859470.962,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564859521.197, "dur": 76.174,
    "args": {
      "External id": 6571,      "device": 2, "stream": 0,      "correlation": 2426, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2426, "pid": 2, "tid": 0, "ts": 7429564859521.197,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564859597.372, "dur": 4.811,
    "args": {
      "External id": 6571,      "device": 2, "stream": 0,      "correlation": 2427, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2427, "pid": 2, "tid": 0, "ts": 7429564859597.372,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859602.184, "dur": 4.048,
    "args": {
      "External id": 6574,      "device": 2, "stream": 0,      "correlation": 2430, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2430, "pid": 2, "tid": 0, "ts": 7429564859602.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564859643.238, "dur": 75.613,
    "args": {
      "External id": 6582,      "device": 2, "stream": 0,      "correlation": 2434, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2434, "pid": 2, "tid": 0, "ts": 7429564859643.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564859718.852, "dur": 5.010,
    "args": {
      "External id": 6582,      "device": 2, "stream": 0,      "correlation": 2435, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2435, "pid": 2, "tid": 0, "ts": 7429564859718.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859723.863, "dur": 4.249,
    "args": {
      "External id": 6585,      "device": 2, "stream": 0,      "correlation": 2438, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2438, "pid": 2, "tid": 0, "ts": 7429564859723.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564859744.631, "dur": 87.520,
    "args": {
      "External id": 6593,      "device": 2, "stream": 0,      "correlation": 2442, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2442, "pid": 2, "tid": 0, "ts": 7429564859744.631,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564859832.152, "dur": 5.051,
    "args": {
      "External id": 6593,      "device": 2, "stream": 0,      "correlation": 2443, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2443, "pid": 2, "tid": 0, "ts": 7429564859832.152,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859837.204, "dur": 4.649,
    "args": {
      "External id": 6596,      "device": 2, "stream": 0,      "correlation": 2444, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2444, "pid": 2, "tid": 0, "ts": 7429564859837.204,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859841.854, "dur": 5.853,
    "args": {
      "External id": 6598,      "device": 2, "stream": 0,      "correlation": 2445, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2445, "pid": 2, "tid": 0, "ts": 7429564859841.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564859847.708, "dur": 7.296,
    "args": {
      "External id": 6601,      "device": 2, "stream": 0,      "correlation": 2446, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2446, "pid": 2, "tid": 0, "ts": 7429564859847.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859855.005, "dur": 5.130,
    "args": {
      "External id": 6602,      "device": 2, "stream": 0,      "correlation": 2447, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2447, "pid": 2, "tid": 0, "ts": 7429564859855.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564859860.136, "dur": 4.570,
    "args": {
      "External id": 6603,      "device": 2, "stream": 0,      "correlation": 2448, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2448, "pid": 2, "tid": 0, "ts": 7429564859860.136,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564859864.707, "dur": 4.690,
    "args": {
      "External id": 6604,      "device": 2, "stream": 0,      "correlation": 2449, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2449, "pid": 2, "tid": 0, "ts": 7429564859864.707,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564859869.398, "dur": 4.088,
    "args": {
      "External id": 6606,      "device": 2, "stream": 0,      "correlation": 2450, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2450, "pid": 2, "tid": 0, "ts": 7429564859869.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564859912.016, "dur": 67.474,
    "args": {
      "External id": 6614,      "device": 2, "stream": 0,      "correlation": 2454, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2454, "pid": 2, "tid": 0, "ts": 7429564859912.016,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564859979.491, "dur": 15.755,
    "args": {
      "External id": 6627,      "device": 2, "stream": 0,      "correlation": 2458, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2458, "pid": 2, "tid": 0, "ts": 7429564859979.491,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564859995.247, "dur": 3.928,
    "args": {
      "External id": 6627,      "device": 2, "stream": 0,      "correlation": 2459, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2459, "pid": 2, "tid": 0, "ts": 7429564859995.247,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564860016.055, "dur": 16.958,
    "args": {
      "External id": 6640,      "device": 2, "stream": 0,      "correlation": 2463, "kind": "Dispatch Kernel",      "grid":     [20480, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2463, "pid": 2, "tid": 0, "ts": 7429564860016.055,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU8_VW4", "pid": 2, "tid": 0,
    "ts": 7429564860033.014, "dur": 3.968,
    "args": {
      "External id": 6640,      "device": 2, "stream": 0,      "correlation": 2464, "kind": "Dispatch Kernel",      "grid":     [3328, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2464, "pid": 2, "tid": 0, "ts": 7429564860033.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860048.289, "dur": 3.407,
    "args": {
      "External id": 6650,      "device": 2, "stream": 0,      "correlation": 2465, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2465, "pid": 2, "tid": 0, "ts": 7429564860048.289,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860076.875, "dur": 2.565,
    "args": {
      "External id": 6655,      "device": 2, "stream": 0,      "correlation": 2466, "kind": "Dispatch Kernel",      "grid":     [104, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2466, "pid": 2, "tid": 0, "ts": 7429564860076.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564860096.800, "dur": 6.575,
    "args": {
      "External id": 6656,      "device": 2, "stream": 0,      "correlation": 2467, "kind": "Dispatch Kernel",      "grid":     [13, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2467, "pid": 2, "tid": 0, "ts": 7429564860096.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860135.369, "dur": 3.006,
    "args": {
      "External id": 6657,      "device": 2, "stream": 0,      "correlation": 2468, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2468, "pid": 2, "tid": 0, "ts": 7429564860135.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860149.923, "dur": 3.166,
    "args": {
      "External id": 6658,      "device": 2, "stream": 0,      "correlation": 2469, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2469, "pid": 2, "tid": 0, "ts": 7429564860149.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860157.941, "dur": 2.124,
    "args": {
      "External id": 6659,      "device": 2, "stream": 0,      "correlation": 2470, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2470, "pid": 2, "tid": 0, "ts": 7429564860157.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::neg_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860172.454, "dur": 2.645,
    "args": {
      "External id": 6664,      "device": 2, "stream": 0,      "correlation": 2471, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2471, "pid": 2, "tid": 0, "ts": 7429564860172.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 4, 64, 64>(at::native::(anonymous namespace)::OpaqueType<4u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata<at::native::(anonymous namespace)::OpaqueType<4u>, unsigned int, 64, 64>, at::native::(anonymous namespace)::TensorSizeStride<unsigned int, 4u>, int, unsigned int)", "pid": 2, "tid": 0,
    "ts": 7429564860182.277, "dur": 5.812,
    "args": {
      "External id": 6665,      "device": 2, "stream": 0,      "correlation": 2472, "kind": "Dispatch Kernel",      "grid":     [4, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2472, "pid": 2, "tid": 0, "ts": 7429564860182.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860192.982, "dur": 2.925,
    "args": {
      "External id": 6666,      "device": 2, "stream": 0,      "correlation": 2473, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2473, "pid": 2, "tid": 0, "ts": 7429564860192.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860201.641, "dur": 3.167,
    "args": {
      "External id": 6667,      "device": 2, "stream": 0,      "correlation": 2474, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2474, "pid": 2, "tid": 0, "ts": 7429564860201.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860264.025, "dur": 3.126,
    "args": {
      "External id": 6684,      "device": 2, "stream": 0,      "correlation": 2475, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2475, "pid": 2, "tid": 0, "ts": 7429564860264.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860296.058, "dur": 4.009,
    "args": {
      "External id": 6702,      "device": 2, "stream": 0,      "correlation": 2476, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2476, "pid": 2, "tid": 0, "ts": 7429564860296.058,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564860346.053, "dur": 6.454,
    "args": {
      "External id": 6715,      "device": 2, "stream": 0,      "correlation": 2480, "kind": "Dispatch Kernel",      "grid":     [8192, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2480, "pid": 2, "tid": 0, "ts": 7429564860346.053,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul> >(int, at::native::AUnaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564860363.052, "dur": 4.490,
    "args": {
      "External id": 6718,      "device": 2, "stream": 0,      "correlation": 2481, "kind": "Dispatch Kernel",      "grid":     [6, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2481, "pid": 2, "tid": 0, "ts": 7429564860363.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::CUDAFunctor_add<float> >(at::TensorIteratorBase&, at::native::CUDAFunctor_add<float> const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860381.615, "dur": 3.647,
    "args": {
      "External id": 6725,      "device": 2, "stream": 0,      "correlation": 2482, "kind": "Dispatch Kernel",      "grid":     [22, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2482, "pid": 2, "tid": 0, "ts": 7429564860381.615,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 4, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429564860398.895, "dur": 3.286,
    "args": {
      "External id": 6728,      "device": 2, "stream": 0,      "correlation": 2483, "kind": "Dispatch Kernel",      "grid":     [26, 1, 1], "block":     [16, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2483, "pid": 2, "tid": 0, "ts": 7429564860398.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA0_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564860454.142, "dur": 5.451,
    "args": {
      "External id": 6740,      "device": 2, "stream": 0,      "correlation": 2487, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2487, "pid": 2, "tid": 0, "ts": 7429564860454.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1} const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860473.546, "dur": 3.888,
    "args": {
      "External id": 6749,      "device": 2, "stream": 0,      "correlation": 2488, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2488, "pid": 2, "tid": 0, "ts": 7429564860473.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564860526.067, "dur": 67.875,
    "args": {
      "External id": 6759,      "device": 2, "stream": 0,      "correlation": 2492, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2492, "pid": 2, "tid": 0, "ts": 7429564860526.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564860593.943, "dur": 4.690,
    "args": {
      "External id": 6762,      "device": 2, "stream": 0,      "correlation": 2493, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2493, "pid": 2, "tid": 0, "ts": 7429564860593.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564860598.634, "dur": 5.211,
    "args": {
      "External id": 6764,      "device": 2, "stream": 0,      "correlation": 2494, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2494, "pid": 2, "tid": 0, "ts": 7429564860598.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564860603.846, "dur": 7.256,
    "args": {
      "External id": 6767,      "device": 2, "stream": 0,      "correlation": 2495, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2495, "pid": 2, "tid": 0, "ts": 7429564860603.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564860611.103, "dur": 4.569,
    "args": {
      "External id": 6768,      "device": 2, "stream": 0,      "correlation": 2496, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2496, "pid": 2, "tid": 0, "ts": 7429564860611.103,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564860615.673, "dur": 4.730,
    "args": {
      "External id": 6769,      "device": 2, "stream": 0,      "correlation": 2497, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2497, "pid": 2, "tid": 0, "ts": 7429564860615.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860620.404, "dur": 5.051,
    "args": {
      "External id": 6770,      "device": 2, "stream": 0,      "correlation": 2498, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2498, "pid": 2, "tid": 0, "ts": 7429564860620.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564860625.456, "dur": 4.128,
    "args": {
      "External id": 6772,      "device": 2, "stream": 0,      "correlation": 2499, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2499, "pid": 2, "tid": 0, "ts": 7429564860625.456,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564860742.084, "dur": 78.339,
    "args": {
      "External id": 6780,      "device": 2, "stream": 0,      "correlation": 2505, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2505, "pid": 2, "tid": 0, "ts": 7429564860742.084,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564860820.424, "dur": 4.930,
    "args": {
      "External id": 6780,      "device": 2, "stream": 0,      "correlation": 2506, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2506, "pid": 2, "tid": 0, "ts": 7429564860820.424,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::silu_kernel(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564860825.355, "dur": 4.169,
    "args": {
      "External id": 6783,      "device": 2, "stream": 0,      "correlation": 2507, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2507, "pid": 2, "tid": 0, "ts": 7429564860825.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT256x16x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT4_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS0_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO0_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564860849.290, "dur": 76.695,
    "args": {
      "External id": 6791,      "device": 2, "stream": 0,      "correlation": 2513, "kind": "Dispatch Kernel",      "grid":     [215040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2513, "pid": 2, "tid": 0, "ts": 7429564860849.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564860925.986, "dur": 5.211,
    "args": {
      "External id": 6791,      "device": 2, "stream": 0,      "correlation": 2514, "kind": "Dispatch Kernel",      "grid":     [46592, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2514, "pid": 2, "tid": 0, "ts": 7429564860925.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul> >(int, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> >, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564860931.198, "dur": 4.891,
    "args": {
      "External id": 6794,      "device": 2, "stream": 0,      "correlation": 2515, "kind": "Dispatch Kernel",      "grid":     [182, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2515, "pid": 2, "tid": 0, "ts": 7429564860931.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x16x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564860936.090, "dur": 87.921,
    "args": {
      "External id": 6802,      "device": 2, "stream": 0,      "correlation": 2519, "kind": "Dispatch Kernel",      "grid":     [245760, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2519, "pid": 2, "tid": 0, "ts": 7429564860936.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429564861024.012, "dur": 4.569,
    "args": {
      "External id": 6802,      "device": 2, "stream": 0,      "correlation": 2520, "kind": "Dispatch Kernel",      "grid":     [13312, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2520, "pid": 2, "tid": 0, "ts": 7429564861024.012,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul> >(int, at::native::CUDAFunctor_add<float>, std::array<char*, 3ul>)", "pid": 2, "tid": 0,
    "ts": 7429564861028.582, "dur": 4.610,
    "args": {
      "External id": 6805,      "device": 2, "stream": 0,      "correlation": 2521, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2521, "pid": 2, "tid": 0, "ts": 7429564861028.582,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::(anonymous namespace)::pow_tensor_scalar_kernel_impl<float, float>(at::TensorIteratorBase&, float)::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564861033.193, "dur": 4.690,
    "args": {
      "External id": 6807,      "device": 2, "stream": 0,      "correlation": 2522, "kind": "Dispatch Kernel",      "grid":     [52, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2522, "pid": 2, "tid": 0, "ts": 7429564861033.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<512, 1, at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4> >(at::native::ReduceOp<float, at::native::MeanOps<float, float, float, float>, unsigned int, float, 4, 4>)", "pid": 2, "tid": 0,
    "ts": 7429564861037.884, "dur": 7.215,
    "args": {
      "External id": 6810,      "device": 2, "stream": 0,      "correlation": 2523, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2523, "pid": 2, "tid": 0, "ts": 7429564861037.884,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul> >(int, at::native::CUDAFunctorOnSelf_add<float>, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564861045.100, "dur": 4.449,
    "args": {
      "External id": 6811,      "device": 2, "stream": 0,      "correlation": 2524, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2524, "pid": 2, "tid": 0, "ts": 7429564861045.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul> >(int, at::native::rsqrt_kernel_cuda(at::TensorIteratorBase&)::{lambda()#2}::operator()() const::{lambda()#2}::operator()() const::{lambda(float)#1}, std::array<char*, 2ul>)", "pid": 2, "tid": 0,
    "ts": 7429564861049.550, "dur": 4.650,
    "args": {
      "External id": 6812,      "device": 2, "stream": 0,      "correlation": 2525, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2525, "pid": 2, "tid": 0, "ts": 7429564861049.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564861054.201, "dur": 4.690,
    "args": {
      "External id": 6813,      "device": 2, "stream": 0,      "correlation": 2526, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2526, "pid": 2, "tid": 0, "ts": 7429564861054.201,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429564861058.892, "dur": 4.048,
    "args": {
      "External id": 6815,      "device": 2, "stream": 0,      "correlation": 2527, "kind": "Dispatch Kernel",      "grid":     [208, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2527, "pid": 2, "tid": 0, "ts": 7429564861058.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x48x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_3_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG64_4_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429564861177.404, "dur": 675.512,
    "args": {
      "External id": 6829,      "device": 2, "stream": 0,      "correlation": 2531, "kind": "Dispatch Kernel",      "grid":     [1539072, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2531, "pid": 2, "tid": 0, "ts": 7429564861177.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU4_VW4", "pid": 2, "tid": 0,
    "ts": 7429564861852.917, "dur": 8.619,
    "args": {
      "External id": 6829,      "device": 2, "stream": 0,      "correlation": 2532, "kind": "Dispatch Kernel",      "grid":     [417024, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2532, "pid": 2, "tid": 0, "ts": 7429564861852.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 537, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 537, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 537, "tid": 0,
    "args": {
      "sort_index": 537
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 0, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 1, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 2, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 3, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 4, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 5, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 6, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 7, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 8, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 9, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 10, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 11, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 12, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 13, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 14, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429564820496.961, "pid": 15, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429564820496.961, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7429564820496.961, "pid": 0, "tid": 0,
    "args": {
      "name": "thread 0 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 0
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7429564820496.961, "pid": 2, "tid": 0,
    "args": {
      "name": "stream 0 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 0
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7429564820496.961, "pid": 537, "tid": 537,
    "args": {
      "name": "thread 537 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 537, "tid": 537,
    "args": {
      "sort_index": 537
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7429564820496.961, "pid": 537, "tid": 537,
    "args": {
      "name": "thread 537 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7429564820496.961, "pid": 537, "tid": 537,
    "args": {
      "sort_index": 537
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 7429564820388.493, "dur": 41484.422,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429564820388.493,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 7429564820388.493
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 7429564877729.521
  }
  ],
  "traceName": "./log_llama/felix-test-job-vllm-3-4dknr_537.1750951162923061071.pt.trace.json",
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1743521598000000000
}